RunPod template · documentation

Gmanski ComfyUI Studio.

The complete README for the pod template: ports and logins, environment variables, every model pack with sizes and sources, volume sizing, first boot and GPU picks.

Deploy on RunPodTemplate overview

A ready-to-run ComfyUI pod: 69 custom node packs, the Gmanski nodes (Song Planner for MiniMax Music 3, Prompt Enhancer, Face Restore, Tile Upscale, Film Grain, Auto Caption), a local Ollama server, and finished workflows for photoreal images, character datasets, upscaling and full-length music.

Deploy with a network volume (200 GB recommended) at /workspace. Without one the pod runs in safe mode and downloads nothing large.

Ports

Port Service Notes
8188 ComfyUI your studio
8888 JupyterLab terminal + files at /workspace; token printed in the pod log banner
8080 FileBrowser drag-and-drop uploads to /workspace; admin + password printed in the pod log banner

The Jupyter token and FileBrowser password are random per pod (set JUPYTER_TOKEN / FILEBROWSER_PASSWORD to pin your own). ComfyUI on 8188 has no login — treat the pod URL as a secret.

Logging in to JupyterLab and FileBrowser

You don't need either to use ComfyUI — they're for the terminal and for moving files.

Find your logins: open the pod in RunPod → Logs (container logs) → scroll to the box printed at the end of startup:

JupyterLab    https://<pod-id>-8888.proxy.runpod.net?token=<token>
FileBrowser   https://<pod-id>-8080.proxy.runpod.net   admin / <password>

Environment variables (all optional)

Variable Default What it does
GMANSKI_PACKS empty Comma-separated model packs to download on boot, e.g. music,zimage. all = everything (needs 250 GB+). Packs can also be added later from the terminal.
CIVITAI_TOKEN empty Civitai API key (civitai.com → Account settings → API Keys). Required for every Civitai-hosted file — zimage, sdxl, flux1, the MiniMax H3 community checkpoints and LoRAs, krea2-intorealism; those files are skipped with skipped-no-token without it. Pre-declared on the template so you can fill it in at deploy time.
HF_TOKEN empty HuggingFace token (huggingface.co → Settings → Access Tokens). Only needed for gated repos (none in the current packs). Pre-declared on the template.
JUPYTER_TOKEN random Pin the JupyterLab token instead of the per-pod random one.
FILEBROWSER_PASSWORD random Pin the FileBrowser admin password instead of the per-pod random one.
GMANSKI_API_KEY empty Gmanski Cloud key for Song Planner provider=gmanski_cloud (Claude Opus 5 / Grok, billed in credits). The local ollama provider needs no key.
GMANSKI_LOWVRAM false Start ComfyUI with --lowvram. Only for 16-24 GB cards that hit out-of-memory on heavy workflows; it moves text encoders to the CPU and breaks MiniMax Music 3.
GMANSKI_ALLOW_NO_VOLUME false Download packs even without a volume (onto the 30 GB container disk; wiped on restart).
GMANSKI_COMFY_EXTRA_ARGS empty Extra ComfyUI command-line flags, appended verbatim.

Model packs

Nothing large downloads until you choose a pack. Set GMANSKI_PACKS on the template, or open the JupyterLab terminal and run:

gmanski-packs list                # packs, sizes, what is already present
gmanski-packs status              # which workflows are ready
gmanski-packs install music       # download one or more packs (resumable)
export CIVITAI_TOKEN=xxxx && gmanski-packs install zimage sdxl

Progress: tail -f /workspace/logs/packs.log. Press R in ComfyUI when a pack finishes.

<!-- gmanski-packs:docs:start -->
Pack What it enables Download Needs
detectors YOLO face/person/eye detectors, SAM, GFPGAN. Installed on every boot (FaceDetailer, Face Restore). 1 GB always installed
music MiniMax Music 3 (full songs from text) + gemma3:27b for Song Planner (Gmanski). 39 GB nothing
minimax-h3 MiniMax H3 shared base (text encoder + video/audio VAEs). Needed by minimax-h3-ref2v. 22 GB nothing
minimax-h3-ref2v MiniMax H3 reference-to-video (turbo): real generated video with a character's identity locked from a reference photo, for Kids Music Video H3 / Music Video H3 / Character Portrait H3 (Gmanski). 48 GB+ VRAM recommended. 23 GB packs: minimax-h3
minimax-h3-flow MiniMax H3 flow (turbo): text-to-video, image-to-video, and first+last-frame-to-video in one model, for Gmanski Flow (Simple)/(Advanced). 48 GB+ VRAM recommended. 23 GB packs: minimax-h3
minimax-h3-image MiniMax H3 photoreal image generator: a LoRA (JonXL, Civitai) that turns the flow checkpoint into a still-photo tool, for Gmanski Flow (Image). 48 GB+ VRAM recommended. 22 GB CIVITAI_TOKEN, packs: minimax-h3
ltx-video-refine LTX-2.5 (22B, INT8): a 2x spatial latent upscale + short refine pass for an existing video, for Video Upscale LTX (Gmanski). Separate model family from MiniMax H3 - no shared base. 48 GB+ VRAM recommended. 40 GB nothing
zimage Z-Image Turbo V1: fast photoreal generation with FaceDetailer and SeedVR2 7B upscale. 24 GB CIVITAI_TOKEN, packs: detectors
upscale Image Upscale SeedVR2: Z-Image Turbo generate + SeedVR2 3B GGUF upscale with Detail Daemon. 24 GB nothing
flux1 Flux 1 Realistic: Flux.1 dev with realism LoRA and upscalers. 34 GB CIVITAI_TOKEN
esrgan-upscale 4xLSDIR pixel-space upscale model, for Video Upscale ESRGAN (Gmanski). Trivial size - shared with flux1 if already installed. 69 MB nothing
sdxl SDXL Img2Img Auto-Prompt: two photoreal SDXL checkpoints (needs the captioning pack for the auto-prompt). 13 GB CIVITAI_TOKEN, packs: captioning, detectors
qwen-edit Qwen Edit 2511 Dataset 30 Images: character dataset generator (bf16, large; 48 GB+ VRAM recommended). 59 GB nothing
captioning Florence2 Folder Captioner and the Auto Caption node (Florence-2 PromptGen + Qwen2.5-VL 7B; JoyCaption optional). 20 GB nothing
training ai-toolkit for the realtime LoRA trainer node (clone + own venv, ~10 GB, several minutes). 10 GB nothing
flux2 Flux 2 Multiple Characters: FLUX.2 [dev] run locally - four reference photos composed into one scene via chained ReferenceLatent conditioning. Was built on the Flux.2 [pro] API node until 2026-09-21, which required a paid comfy.org login and billed per image; these local weights replace it entirely, no account and no per-image cost. 40 GB+ VRAM recommended. 50 GB nothing
minimax-h3-dasiwa DaSiWa MiniMax H3 Hybrid v4 (Civitai community fine-tune/merge of ref2va + optimised fl2va, w6a8) - an alternate checkpoint for Gmanski Flow (DaSiWa) and Character Portrait H3 (DaSiWa), not a default swap. Optional: the 4/8-step Turbo variant (--with-optional). Hybrid V1 int8/int4 stay as optional files for older workflows. 16 GB CIVITAI_TOKEN, packs: minimax-h3
minimax-h3-remix MiniMax H3 Remix int8_convrot v0.6 fixed (Civitai community dual-H3 merge with LoRAs statically fused in, ships its own paired text encoder) - an alternate checkpoint, not a default swap. Tensor-header-verified to share most of MiniMax H3's underlying structure for both flow and reference-to-video tasks (89.9% match). 2026-11-07: v0.8 HSA turbo (pruned int8) is the default checkpoint; the v0.6 DiT stays as an optional file. 47 GB CIVITAI_TOKEN, packs: minimax-h3
minimax-h3-motion Better Human Motion (Civitai, MiniMax H3 version, 344 thumbs-up): a motion-quality LoRA for natural, consistent body movement in generated video. 303 MB CIVITAI_TOKEN, packs: minimax-h3-flow
krea2 Krea 2 RAW (official Comfy-Org checkpoint) + turbo LoRA + Qwen3-VL 4B text encoder + Wan 2.1 FP32 VAE, for Gmanski Krea 2. Sourced from a real, tested community workflow (Lonecat's Krea 2 simple gen), not guessed - node graph and exact filenames verified against it directly. That workflow also relied on a 'filter bypass' LoRA (Civitai model 2728234) for prompt adherence; dropped 2026-09-20 because it is early-access/paid on Civitai (the download endpoint redirects to the model page even with a valid token), so this pack runs on the turbo LoRA alone and prompt adherence may be lower than the reference workflow's. 34 GB nothing
krea2-intorealism IntoRealism Krea 2 v5.0 (int8, Civitai community fine-tune; v4.0's fp8 weights peppered skin with dark specks - see the template's note) - an alternate checkpoint for Gmanski Krea 2 (IntoRealism), not a default swap. Shares the base krea2 pack's turbo LoRA, text encoder and VAE. 13 GB CIVITAI_TOKEN, packs: krea2
minimax-h3-refine MiniMax H3 native refine kit (ComfyUI >= 0.44): the 3D latent upscaler (Load Latent Upscale Model + MiniMax H3 Latent Upscale With Model, Apache-2.0, 0.7 GB) and NVIDIA's Sol-H3-VAE fast decoder (int8; fp16 optional). Lets Gmanski Flow (Long) re-sample a finished clip at 1.5x inside the same context windows instead of a VAE round trip. Required by minimax-h3-upscaling. 3 GB packs: minimax-h3
minimax-h3-upscaling MiniMax H3 Upscaling: 2x latent upscale before decode via a learned neural upscaler (LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler custom node, not naive interpolation - avoids the ghosting/double-image artifacts bilinear/bicubic upscaling causes on H3 latents). Uses its own bf16 fl2va checkpoint, int8_convrot text encoder and 768p-tuned turbo LoRAs - deliberately NOT reused from the existing int8-pruned minimax-h3-flow pack, since this workflow set was verified against this exact bf16/int8_convrot combination and mixing precisions with an untested pairing is exactly the kind of thing that caused a real shape-mismatch crash earlier in this project. Reuses the existing ref2va checkpoint, video VAE and audio VAE unchanged (verified identical filenames). For the 5 MiniMax Upscaling (Gmanski) video templates (T2V/I2V/R2V, Auto/Custom Prompt). 48 GB+ VRAM recommended. 2026-11-07: the latent upscaler and the Sol-H3-VAE fast decoder moved to the small minimax-h3-refine pack (required here). 97 GB packs: minimax-h3, minimax-h3-refine
minimax-h3-quality Two optional MiniMax H3 quality LoRAs (Civitai). Natural Face & Speech (AdaptiveVision) adds face-muscle dynamics and clearer English speech - H3 generates its own audio, and nothing else shipped here improves talking-head quality; recommended weight 0.4-0.8, and the author advises a fixed seed plus a consistent opening word to stop the voice drifting. Authentic Cinematic Texture (TuTu_1018) is a film-look pass, trigger word DY, strength 0.7 dropping to 0.5 on high-motion shots. Both stack onto any H3 pipeline. 594 MB CIVITAI_TOKEN, packs: minimax-h3
minimax-h3-turbo4 lightx2v's 4-step turbo LoRA v1.2 768p (via tsolful's Minimax H3 Turbo Loras collection) - same lineage as the 8-step turbo already used by the flow templates, so it drops straight in at roughly twice the speed. Author's guidance: 5-8 steps, er_sde or res_multistep, simple scheduler. An alternative to the 8-step, not an addition - do not stack both. 2 GB CIVITAI_TOKEN, packs: minimax-h3
krea2-checkpoints-realism Three community Krea 2 realism fine-tunes as alternate checkpoints (Civitai, 2026-09-29 sweep of the top Krea 2 models): Realism by Stable Yogi v3.0 INT8 (portraits/lifestyle, 8 steps CFG 1), Muse by Stable Yogi v3.5 Int8 Extended (fashion/editorial, 12-16 steps CFG 1.5-2) and GonzaLomo v4.0 (photoreal). Picked per picture by the Krea 2 LoRA Router's checkpoint routing in Bulk Img2Img Character Krea 2 Dynamic LoRA; each runs WITHOUT the turbo LoRA (they are turbo-tuned themselves). 2026-11-07: Muse bumped to V4.0 Extended; CyberRealistic Krea 2 V4.0 added as a manual route. 51 GB CIVITAI_TOKEN, packs: krea2
krea2-body-sliders Body-shape LoRAs for Krea 2 (Civitai): Loraholic's figure and age sliders (signed strengths, no trigger - the router picks the sign from the caption; the age slider is only ever routed toward OLDER automatically), plus three figure-shape LoRAs. 896 MB CIVITAI_TOKEN, packs: krea2
krea2-realism-loras Style / realism LoRAs for Krea 2 (Civitai): Famegrid (trigger Famegrid, 0.3-0.6 with a character LoRA), UltraReal v2 Pro (0.6-0.8, no trigger), BloomGirls UltraRealism (bl00m), Krea2-realism V2 (1.0-2.0), Detail slider, Realistic slider. One at a time: the router never stacks two styles and never picks one from a caption on its own - choose it on the node or name it in Extra words. 4 GB CIVITAI_TOKEN, packs: krea2
krea2-v2 The realism LoRA Gmanski Krea 2 v2 adds to the base krea2 pack: Realism Engine v3.1 (Hugging Face, no token). 2 GB packs: krea2
krea2-clothing Clothing LoRAs for Krea 2 (Civitai): MicroSkirt (trigger microskirt) and the Kukuro outfit set. 446 MB CIVITAI_TOKEN, packs: krea2
krea2-identity-edit Krea 2 Identity Edit V2 + General Edit V2 (Civitai, conrad_locke, 2026-10-16): instruction-based, identity-preserving editing - put the person from a reference PHOTO into another picture's scene with no LoRA training, plus plain-language edits. Needs the ComfyUI-Krea2Edit node pack >= 2.0.0 (installed in the image). v1.2 ships as an optional file for older workflows. For Bulk Img2Img Reference Krea 2 (Gmanski). 4 GB CIVITAI_TOKEN, packs: krea2
krea2-intorealism-flash IntoRealism Krea 2 Flash V3 (int8, 4 steps) - the fast distilled sibling of the krea2-intorealism pack's v5.0. Alternate checkpoint: manual pick in the router (checkpoint_mode intorealism_flash) or the Gmanski Krea 2 (IntoRealism Flash) template; shares krea2's text encoder and VAE. 13 GB CIVITAI_TOKEN, packs: krea2
minimax-h3-controlnet MiniMax H3 Fun-ControlNet-Union 2.0 (Alibaba PAI, int8 patch) for pose/depth/edge-driven video: DWPose skeletons from a driver clip make people and pets do choreography that prompt-only H3 smears. Loads on the flow checkpoint via ModelPatchLoader. v1 is optional for A/B. 5 GB packs: minimax-h3-flow
minimax-h3-turbo-alt Alternative MiniMax H3 turbo/distill LoRAs for the 8-step A/B against the lightx2v v1.0 default in minimax-h3-flow / -ref2v: Pika H3 Flex (fl2va + ref2va, 2026-11-06) and larryvrh v4 step600 EMA. Nothing ships wired to these until the quality matrix picks one. 5 GB CIVITAI_TOKEN, packs: minimax-h3
minimax-h3-character-swap Character Video Swap: replace the performer in a guide video with a reference photo's person while keeping the choreography, camera and background (BFS H3 Character Swap v1.1 LoRA on the ref2va checkpoint). For Gmanski Character Video Swap (H3). 92 MB CIVITAI_TOKEN, packs: minimax-h3-ref2v
minimax-h3-pimage P-H3-Image (Pruna AI, from MiniMax H3): 8-step text-to-image and multi-reference editing (up to 8 pictures, 8 MP) with no CFG. Ships as a 2 GB LoRA on the minimax-h3-flow checkpoint plus its own text encoder and VAE (int8), the way ComfyUI's official template runs it. Optional: prompt upsampler, merged 21 GB DiT. For Gmanski H3 Image (P-H3) and H3 Image Edit (P-H3). ComfyUI >= 0.44 only. 10 GB packs: minimax-h3-flow
moss MOSS-Image (OpenMOSS, Apache-2.0, 2026-10-16) - the 5B text-to-image model that topped Civitai in November 2026: accurate in-image text, JSON layout boxes, 4K-token prompts, 2K output. SFT checkpoint + Qwen3.5-9B int8 encoder + FLUX.2 VAE (shared with the flux2 pack). Optional: Base (for LoRA training), official edit LoRA, colour/depth control. For Gmanski MOSS and Gmanski MOSS Layout. ComfyUI >= 0.42. 15 GB nothing
moss-realism Photoreal MOSS-Image community checkpoints (int8): Realism by Stable Yogi Moss V1.5 (8 steps) and Moody Moss Mix V4.1 (25 steps, CFG 3). Alternate checkpoints for Gmanski MOSS, not default swaps. 10 GB CIVITAI_TOKEN, packs: moss
moss-turbo Step-distillation LoRAs for MOSS-Image: 8-step DMD2 (recommended) and 4-step SFT v2. Optional speed-ups for the Gmanski MOSS templates (bypassed by default). 2 GB CIVITAI_TOKEN, packs: moss
ace-step ACE-Step 1.5 XL (Apache-2.0) - a second, licence-clean music engine next to MiniMax Music 3: lyrics + style caption to a full song in 8 steps (Turbo), with the 0.6B + 4B Qwen text encoders and the 1.5 VAE from the official ComfyUI template. Optional: XL SFT for cover/repaint. For Gmanski Music (ACE-Step). 20 GB nothing

Templates and the packs they need:

Template Packs Bring your own
Z-Image Turbo V1 (Gmanski) zimage
Image Upscale SeedVR2 (Gmanski) upscale ZIT_detailed_eyes_V2_lora.safetensors
MiniMax Music 3 (Gmanski) music
Kids Music Video (Gmanski) music, zimage
Kids Music Video H3 (Gmanski) music, minimax-h3-ref2v
Music Video (Gmanski) music, zimage
Music Video H3 (Gmanski) music, minimax-h3-ref2v
Gmanski Flow (Simple) minimax-h3-flow
Gmanski Flow (Advanced) minimax-h3-flow
Gmanski Flow (Image) minimax-h3-image
Gmanski Flow (Image Advanced) minimax-h3-image
Character Portrait H3 (Gmanski) minimax-h3-ref2v
Character Ensemble H3 (Gmanski) minimax-h3-ref2v
Character Ensemble H3 Video (Gmanski) minimax-h3-ref2v
Video Upscale LTX (Gmanski) ltx-video-refine
Video Upscale ESRGAN (Gmanski) esrgan-upscale
Register a Character (Gmanski) none
Flux 1 Realistic (Gmanski) flux1
Qwen Edit 2511 Dataset 30 Images (Gmanski) qwen-edit
SDXL Img2Img Auto-Prompt (Gmanski) sdxl character_lora.safetensors
Florence2 Folder Captioner (Gmanski) captioning
Flux 2 Multiple Characters (Gmanski) flux2
Gmanski Flow (DaSiWa) minimax-h3-dasiwa
Character Portrait H3 (DaSiWa) (Gmanski) minimax-h3-dasiwa
Gmanski Flow (Remix) minimax-h3-remix
Character Portrait H3 (Remix) (Gmanski) minimax-h3-remix
Gmanski Flow (Continuation) minimax-h3-flow
Gmanski Flow (Extend) minimax-h3-flow, minimax-h3-ref2v
Video Frame Interpolation (Gmanski) none rife49.pth
MiniMax Upscaling (T2V Auto Prompt) (Gmanski) minimax-h3-upscaling
MiniMax Upscaling (T2V Custom Prompt) (Gmanski) minimax-h3-upscaling
MiniMax Upscaling (I2V Auto Prompt) (Gmanski) minimax-h3-upscaling
MiniMax Upscaling (I2V Custom Prompt) (Gmanski) minimax-h3-upscaling
MiniMax Upscaling (R2V Auto Prompt) (Gmanski) minimax-h3-upscaling, minimax-h3-ref2v
Gmanski Krea 2 krea2
Random Influencer Face Generator Krea 2 (Gmanski) krea2
Gmanski Krea 2 v2 krea2-v2
Gmanski Krea 2 (IntoRealism) krea2-intorealism
Video Upscale LTX Long (Gmanski) ltx-video-refine
Bulk Img2Img Character Krea 2 (Gmanski) krea2, captioning, detectors <the character's Krea 2 LoRA, from the library's face_lora_krea2>
Bulk Img2Img Character Krea 2 Dynamic LoRA (Gmanski) krea2, krea2-intorealism, krea2-body-sliders, krea2-realism-loras, krea2-clothing, krea2-checkpoints-realism, captioning, detectors <the character's Krea 2 LoRA, from the library's face_lora_krea2>
Bulk Img2Img Reference Krea 2 (Gmanski) krea2, krea2-identity-edit, krea2-body-sliders, krea2-realism-loras, krea2-clothing, captioning, detectors
Gmanski Flow (Long) minimax-h3-flow, minimax-h3-refine
Gmanski H3 Image (P-H3) minimax-h3-flow, minimax-h3-pimage
Gmanski H3 Image Edit (P-H3) minimax-h3-flow, minimax-h3-pimage
Gmanski MOSS moss, moss-turbo
Gmanski MOSS (Layout) moss, moss-turbo
Gmanski Music (ACE-Step) ace-step
Character Video Swap H3 (Gmanski) minimax-h3-ref2v, minimax-h3-character-swap
Gmanski Krea 2 (IntoRealism Flash) krea2, krea2-intorealism-flash

Files per pack (destination under /workspace/models/):

detectors

File Destination Source Size
face_yolov8m.pt ultralytics/bbox/face_yolov8m.pt HF Bingsu/adetailer 53 MB
face_yolov8n.pt ultralytics/bbox/face_yolov8n.pt HF Bingsu/adetailer 7 MB
person_yolov8m-seg.pt ultralytics/segm/person_yolov8m-seg.pt HF Bingsu/adetailer 53 MB
sam_vit_b_01ec64.pth sams/sam_vit_b_01ec64.pth dl.fbaipublicfiles.com 369 MB
PitEyeDetailer-v2-seg.pt ultralytics/bbox/PitEyeDetailer-v2-seg.pt HF ashllay/YOLO_Models 7 MB
Nipplies.pt (optional) ultralytics/bbox/Nipplies.pt pixeldrain.com 7 MB
GFPGANv1.4.pth face_restore/GFPGANv1.4.pth github.com 338 MB

music

File Destination Source Size
minimax_music3_dit_fp16.safetensors diffusion_models/minimax_music3_dit_fp16.safetensors HF Comfy-Org/MiniMax-Music-3 4.9 GB
minimax_music3_text_encoder_pruned_bf16.safetensors text_encoders/minimax_music3_text_encoder_pruned_bf16.safetensors HF Comfy-Org/MiniMax-Music-3 16.7 GB
minimax_music3_dav.safetensors vae/minimax_music3_dav.safetensors HF Comfy-Org/MiniMax-Music-3 225 MB
Ollama gemma3:27b /workspace/ollama-models ollama.com 17 GB

minimax-h3

File Destination Source Size
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors HF Comfy-Org/MiniMax-H3 15.7 GB
minimax_h3_video_vae_fp16.safetensors vae/minimax_h3_video_vae_fp16.safetensors HF Comfy-Org/MiniMax-H3 5.2 GB
minimax_h3_audio_vae_fp32.safetensors vae/minimax_h3_audio_vae_fp32.safetensors HF Comfy-Org/MiniMax-H3 625 MB
taeh3.safetensors (optional) vae_approx/taeh3.safetensors HF t8star/Taeh3-Comfy 24 MB

minimax-h3-ref2v

File Destination Source Size
minimax_h3_ref2va_pruned_int8_convrot.safetensors diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors HF Comfy-Org/MiniMax-H3 21.0 GB
minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors loras/minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors HF lightx2v/Minimax-h3-Turbo 2.0 GB

minimax-h3-flow

File Destination Source Size
minimax_h3_fl2va_pruned_int8_convrot.safetensors diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors HF Comfy-Org/MiniMax-H3 21.0 GB
minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors loras/minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors HF lightx2v/Minimax-h3-Turbo 2.0 GB

minimax-h3-image

File Destination Source Size
minimax_h3_fl2va_pruned_int8_convrot.safetensors diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors HF Comfy-Org/MiniMax-H3 21.0 GB
h3_photoreal_ph0t0r34l_1024-step00003780.safetensors loras/h3_photoreal_ph0t0r34l_1024-step00003780.safetensors Civitai version 3260165 (token) 614 MB

ltx-video-refine

File Destination Source Size
ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors diffusion_models/ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors HF Lightricks/LTX-2.5 21.5 GB
gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors text_encoders/gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors HF Lightricks/LTX-2.5 15.4 GB
ltx-2.5-video-vae-bf16.safetensors vae/ltx-2.5-video-vae-bf16.safetensors HF Lightricks/LTX-2.5 1.5 GB
ltx-2.5-audio-vae-bf16.safetensors vae/ltx-2.5-audio-vae-bf16.safetensors HF Lightricks/LTX-2.5 379 MB
ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors HF Lightricks/LTX-2.5 1.0 GB

zimage

File Destination Source Size
zimageTurboByStable_2601Fp8.safetensors checkpoints/zimageTurboByStable_2601Fp8.safetensors Civitai version 2576181 (token) 5.7 GB
qwen_3_4b.safetensors z_image/qwen_3_4b.safetensors HF Comfy-Org/z_image_turbo 8.0 GB
ae.safetensors ae.safetensors HF Comfy-Org/z_image_turbo 348 MB
seedvr2_ema_7b_sharp_fp8_e4m3fn_mixed_block35_fp16.safetensors SEEDVR2/seedvr2_ema_7b_sharp_fp8_e4m3fn_mixed_block35_fp16.safetensors HF AInVFX/SeedVR2_comfyUI 8.5 GB
ema_vae_fp16.safetensors SEEDVR2/ema_vae_fp16.safetensors HF numz/SeedVR2_comfyUI 512 MB
Eyes.pt ultralytics/bbox/Eyes.pt HF Tenofas/ComfyUI 24 MB

upscale

File Destination Source Size
seedvr2_ema_3b-Q4_K_M.gguf SEEDVR2/seedvr2_ema_3b-Q4_K_M.gguf HF AInVFX/SeedVR2_comfyUI 2.0 GB
ema_vae_fp16.safetensors SEEDVR2/ema_vae_fp16.safetensors HF numz/SeedVR2_comfyUI 512 MB
z_image_turbo_bf16.safetensors z_image/z_image_turbo_bf16.safetensors HF Comfy-Org/z_image_turbo 12.3 GB
qwen_3_4b.safetensors z_image/qwen_3_4b.safetensors HF Comfy-Org/z_image_turbo 8.0 GB
ae.safetensors ae.safetensors HF Comfy-Org/z_image_turbo 348 MB

flux1

File Destination Source Size
flux1-dev.safetensors flux1-dev.safetensors HF Comfy-Org/flux1-dev 23.8 GB
t5xxl_fp16.safetensors text_encoders/t5xxl_fp16.safetensors HF comfyanonymous/flux_text_encoders 9.8 GB
clip_l.safetensors text_encoders/clip_l.safetensors HF comfyanonymous/flux_text_encoders 256 MB
ae.safetensors ae.safetensors HF Comfy-Org/z_image_turbo 348 MB
4xLSDIR.pth upscale/4xLSDIR.pth HF LS110824/upscale 69 MB
1xSkinContrast-SuperUltraCompact.pth upscale/1xSkinContrast-SuperUltraCompact.pth HF reezus/comfyui-models 1 MB
Samsung_UltraReal.safetensors loras/Samsung_UltraReal.safetensors Civitai version 1755780 (token) 154 MB

esrgan-upscale

File Destination Source Size
4xLSDIR.pth upscale/4xLSDIR.pth HF LS110824/upscale 69 MB

sdxl

File Destination Source Size
ultraRealisticByStable_v20FP16.safetensors SDXL/ultraRealisticByStable_v20FP16.safetensors Civitai version 2285644 (token) 6.5 GB
analogMadnessSDXL_x15.safetensors SDXL/analogMadnessSDXL_x15.safetensors Civitai version 2207703 (token) 6.5 GB

qwen-edit

File Destination Source Size
qwen_image_edit_2511_bf16.safetensors diffusion_models/qwen_image_edit_2511_bf16.safetensors HF Comfy-Org/Qwen-Image-Edit_ComfyUI 40.9 GB
qwen_2.5_vl_7b.safetensors clip/qwen_2.5_vl_7b.safetensors HF Comfy-Org/Qwen-Image_ComfyUI 16.6 GB
qwen_image_vae.safetensors vae/qwen_image_vae.safetensors HF Comfy-Org/Qwen-Image_ComfyUI 256 MB
qwen-image-edit-2511-multiple-angles-lora.safetensors loras/qwen-image-edit/qwen-image-edit-2511-multiple-angles-lora.safetensors HF fal/Qwen-Image-Edit-2511-Multiple-Angles-LoRA 307 MB
Qwen-Image-Edit-2511-Lightning-8steps-V1.0-bf16.safetensors loras/qwen-image-edit/Qwen-Image-Edit-2511-Lightning-8steps-V1.0-bf16.safetensors HF lightx2v/Qwen-Image-Edit-2511-Lightning 870 MB

captioning

File Destination Source Size
Florence-2-large-PromptGen-v2.0 LLM/Florence-2-large-PromptGen-v2.0 HF snapshot MiaoshouAI/Florence-2-large-PromptGen-v2.0 3.3 GB
Qwen/Qwen2.5-VL-7B-Instruct HF cache HF snapshot Qwen/Qwen2.5-VL-7B-Instruct 16.6 GB
fancyfeast/llama-joycaption-beta-one-hf-llava (optional) HF cache HF snapshot fancyfeast/llama-joycaption-beta-one-hf-llava 17.0 GB

training

File Destination Source Size
setup script /opt/gmanski/setup_ai_toolkit.sh GitHub ostris/ai-toolkit ~10 GB

flux2

File Destination Source Size
flux2_dev_fp8mixed.safetensors unet/flux2_dev_fp8mixed.safetensors HF Comfy-Org/flux2-dev 33.0 GB
mistral_3_small_flux2_fp8.safetensors clip/mistral_3_small_flux2_fp8.safetensors HF Comfy-Org/flux2-dev 16.8 GB
flux2-vae.safetensors vae/flux2-vae.safetensors HF Comfy-Org/flux2-dev 321 MB

minimax-h3-dasiwa

File Destination Source Size
dasiwa_minimax_h3_ref2va_v4_pruned_w6a8.safetensors diffusion_models/dasiwa_minimax_h3_ref2va_v4_pruned_w6a8.safetensors Civitai version 3391862 (token) 16.4 GB
dasiwa_minimax_h3_ref2va_v4_pruned_turbo_w6a8.safetensors (optional) diffusion_models/dasiwa_minimax_h3_ref2va_v4_pruned_turbo_w6a8.safetensors Civitai version 3392061 (token) 16.4 GB
DasiwaMinimaxH3_dasiwaHybridV1_3156811.safetensors (optional) diffusion_models/DasiwaMinimaxH3_dasiwaHybridV1_3156811.safetensors Civitai version 3251526 (token) 19.5 GB
DasiwaMinimaxH3_dasiwaHybridV1_3177797.safetensors (optional) diffusion_models/DasiwaMinimaxH3_dasiwaHybridV1_3177797.safetensors Civitai version 3251526 (token) 11.7 GB

minimax-h3-remix

File Destination Source Size
FeiHou_MiniMax-H3_Remix_HSA_turbo_v0.8_int8_convrotHQ_pruned.safetensors diffusion_models/FeiHou_MiniMax-H3_Remix_HSA_turbo_v0.8_int8_convrotHQ_pruned.safetensors Civitai version 3423265 (token) 22.7 GB
minimaxH3Remix_int8ConvrotV06Fixed.safetensors (optional) diffusion_models/minimaxH3Remix_int8ConvrotV06Fixed.safetensors Civitai version 3254025 (token) 31.7 GB
minimaxH3Remix_int8ConvrotV06Fixed_txt.safetensors text_encoders/minimaxH3Remix_int8ConvrotV06Fixed_txt.safetensors Civitai version 3254025 (token) 24.6 GB

minimax-h3-motion

File Destination Source Size
mvmt_h3_lora_v1_500.safetensors loras/mvmt_h3_lora_v1_500.safetensors Civitai version 3256084 (token) 303 MB

krea2

File Destination Source Size
krea2_raw_bf16.safetensors diffusion_models/krea2_raw_bf16.safetensors HF Comfy-Org/Krea-2 24.5 GB
krea2_turbo_lora_rank_64_bf16.safetensors loras/krea2_turbo_lora_rank_64_bf16.safetensors HF Comfy-Org/Krea-2 451 MB
qwen3vl_4b_bf16.safetensors text_encoders/qwen3vl_4b_bf16.safetensors HF Comfy-Org/Krea-2 8.3 GB
wan21_vae_fp32.safetensors vae/wan21_vae_fp32.safetensors HF Kijai/WanVideo_comfy 492 MB

krea2-intorealism

File Destination Source Size
intorealismKrea2_v50.safetensors diffusion_models/intorealismKrea2_v50.safetensors Civitai version 3328673 (token) 12.7 GB

minimax-h3-refine

File Destination Source Size
minimax_h3_latent_upscaler_3d_fp16.safetensors latent_upscale_models/minimax_h3_latent_upscaler_3d_fp16.safetensors HF LBH-123-AI/Minimax_h3_latent_Upscaler 707 MB
minimax_h3_sol_vae_int8_convrot.safetensors vae/minimax_h3_sol_vae_int8_convrot.safetensors HF Comfy-Org/MiniMax-H3 2.0 GB
minimax_h3_sol_vae_fp16.safetensors (optional) vae/minimax_h3_sol_vae_fp16.safetensors HF Comfy-Org/MiniMax-H3 3.6 GB

minimax-h3-upscaling

File Destination Source Size
qwen3vl_32b_minimax_h3_int8_convrot.safetensors text_encoders/qwen3vl_32b_minimax_h3_int8_convrot.safetensors HF Comfy-Org/MiniMax-H3 27.1 GB
minimax_h3_fl2va_bf16.safetensors diffusion_models/minimax_h3_fl2va_bf16.safetensors HF Comfy-Org/MiniMax-H3 66.3 GB
minimax_h3_fl2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors loras/minimax_h3_fl2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors HF lightx2v/Minimax-h3-Turbo 2.0 GB
minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors loras/minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors HF lightx2v/Minimax-h3-Turbo 2.0 GB

minimax-h3-quality

File Destination Source Size
nafasp_h3_lora_v2_500.safetensors loras/nafasp_h3_lora_v2_500.safetensors Civitai version 3339123 (token) 297 MB
minimax_h3_authentic_cinematic_texture.safetensors loras/minimax_h3_authentic_cinematic_texture.safetensors Civitai version 3267949 (token) 297 MB

minimax-h3-turbo4

File Destination Source Size
minimax_h3_fl2v_turbo_4step_v1.2_768p_comfyui_bf16.safetensors loras/minimax_h3_fl2v_turbo_4step_v1.2_768p_comfyui_bf16.safetensors Civitai version 3297123 (token) 1.8 GB

krea2-checkpoints-realism

File Destination Source Size
realismByStableYogi_v30INT8FP8Extended_3215326.safetensors diffusion_models/realismByStableYogi_v30INT8FP8Extended_3215326.safetensors Civitai version 3329215 (token) 12.6 GB
museByStableYogi_v40Int8Extended.safetensors diffusion_models/museByStableYogi_v40Int8Extended.safetensors Civitai version 3425119 (token) 13.2 GB
gonzalomoKrea2_v40_3128457.safetensors diffusion_models/gonzalomoKrea2_v40_3128457.safetensors Civitai version 3245099 (token) 11.9 GB
cyberrealisticKrea2_v40Int8.safetensors diffusion_models/cyberrealisticKrea2_v40Int8.safetensors Civitai version 3357068 (token) 13.2 GB

krea2-body-sliders

File Destination Source Size
breast_size_v2_krea2_loraholic.safetensors loras/breast_size_v2_krea2_loraholic.safetensors Civitai version 3131773 (token) 34 MB
ass_v2_krea2_loraholic.safetensors loras/ass_v2_krea2_loraholic.safetensors Civitai version 3207249 (token) 48 MB
pubes_krea2_loraholic.safetensors loras/pubes_krea2_loraholic.safetensors Civitai version 3136749 (token) 13 MB
age_krea2_loraholic.safetensors loras/age_krea2_loraholic.safetensors Civitai version 3067659 (token) 7 MB
pawg_krea2.safetensors loras/pawg_krea2.safetensors Civitai version 3071384 (token) 109 MB
AA_NM_Fitbutt_KREA2.safetensors loras/AA_NM_Fitbutt_KREA2.safetensors Civitai version 3352041 (token) 436 MB
TeeKaysTittyTime_GodTierTits_V1-1_Kr2.safetensors loras/TeeKaysTittyTime_GodTierTits_V1-1_Kr2.safetensors Civitai version 3312813 (token) 218 MB

krea2-realism-loras

File Destination Source Size
famegrid_spicy_3167060.safetensors loras/famegrid_spicy_3167060.safetensors Civitai version 3278885 (token) 1.5 GB
ultra_real_krea2_v2.safetensors loras/ultra_real_krea2_v2.safetensors Civitai version 3182356 (token) 218 MB
krea2-bloomgirls-realism-step00004000.safetensors loras/krea2-bloomgirls-realism-step00004000.safetensors Civitai version 3191538 (token) 447 MB
Krea2-realism-V2.safetensors loras/Krea2-realism-V2.safetensors Civitai version 3090634 (token) 1.5 GB
detail_slider_krea2_loraholic.safetensors loras/detail_slider_krea2_loraholic.safetensors Civitai version 3121642 (token) 20 MB
real_3d_krea2_loraholic.safetensors loras/real_3d_krea2_loraholic.safetensors Civitai version 3066073 (token) 7 MB

krea2-v2

File Destination Source Size
realism_engine_krea2_v3.1.safetensors loras/realism_engine_krea2_v3.1.safetensors HF RazzzHF/realism_engine_krea2 1.6 GB

krea2-clothing

File Destination Source Size
Micro_Skirt_K1_epoch_10.safetensors loras/Micro_Skirt_K1_epoch_10.safetensors Civitai version 3339709 (token) 218 MB
Kukuro_epoch_10.safetensors loras/Kukuro_epoch_10.safetensors Civitai version 3142469 (token) 218 MB

krea2-identity-edit

File Destination Source Size
krea2_identity_edit_v2.safetensors loras/krea2_identity_edit_v2.safetensors Civitai version 3409322 (token) 1.8 GB
krea2_general_edit_v2.safetensors loras/krea2_general_edit_v2.safetensors Civitai version 3409322 (token) 1.8 GB
krea2_identity_edit_v1_2.safetensors (optional) loras/krea2_identity_edit_v1_2.safetensors Civitai version 3139172 (token) 1.7 GB

krea2-intorealism-flash

File Destination Source Size
intorealismKrea2_flashV3Int8.safetensors diffusion_models/intorealismKrea2_flashV3Int8.safetensors Civitai version 3446425 (token) 13.2 GB

minimax-h3-controlnet

File Destination Source Size
minimax_h3_fun_controlnet_union_2.0_pruned_int8_convrot.safetensors model_patches/minimax_h3_fun_controlnet_union_2.0_pruned_int8_convrot.safetensors HF Comfy-Org/MiniMax-H3 4.5 GB
minimax_h3_fun_controlnet_union_pruned_int8_convrot.safetensors (optional) model_patches/minimax_h3_fun_controlnet_union_pruned_int8_convrot.safetensors HF Comfy-Org/MiniMax-H3 2.3 GB

minimax-h3-turbo-alt

File Destination Source Size
pika_h3_flex_fl2va_v1_bf16.safetensors loras/pika_h3_flex_fl2va_v1_bf16.safetensors HF Comfy-Org/MiniMax-H3 2.0 GB
pika_h3_flex_ref2va_v1_bf16.safetensors loras/pika_h3_flex_ref2va_v1_bf16.safetensors HF Comfy-Org/MiniMax-H3 2.0 GB
minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors loras/minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors Civitai version 3202732 (token) 621 MB

minimax-h3-character-swap

File Destination Source Size
minimax_h3_character_swap_v1.1_ref2va_fro.safetensors loras/minimax_h3_character_swap_v1.1_ref2va_fro.safetensors Civitai version 3412620 (token) 81 MB

minimax-h3-pimage

File Destination Source Size
p-h3-image_lora_bf16.safetensors loras/p-h3-image_lora_bf16.safetensors HF PrunaAI/P-H3-Image-Comfy 2.0 GB
p-h3-image_text_encoder_int8_convrot.safetensors text_encoders/p-h3-image_text_encoder_int8_convrot.safetensors HF PrunaAI/P-H3-Image-Comfy 5.1 GB
p-h3-image_vae_int8_convrot.safetensors vae/p-h3-image_vae_int8_convrot.safetensors HF PrunaAI/P-H3-Image-Comfy 3.1 GB
p-h3-image_prompt_upsampler_int8_convrot.safetensors (optional) text_encoders/p-h3-image_prompt_upsampler_int8_convrot.safetensors HF PrunaAI/P-H3-Image-Comfy 9.8 GB
p-h3-image_int8_convrot.safetensors (optional) diffusion_models/p-h3-image_int8_convrot.safetensors HF PrunaAI/P-H3-Image-Comfy 21.2 GB

moss

File Destination Source Size
moss_image_sft_int8_convrot.safetensors diffusion_models/moss_image_sft_int8_convrot.safetensors HF Comfy-Org/MOSS-Image 5.4 GB
qwen3.5_9b_int8_convrot.safetensors text_encoders/qwen3.5_9b_int8_convrot.safetensors HF Comfy-Org/MOSS-Image 9.8 GB
flux2-vae.safetensors vae/flux2-vae.safetensors HF Comfy-Org/flux2-dev 321 MB
moss_image_base_int8_convrot.safetensors (optional) diffusion_models/moss_image_base_int8_convrot.safetensors HF Comfy-Org/MOSS-Image 5.4 GB
moss_image_edit_lora_rank_64_bf16.safetensors (optional) loras/moss_image_edit_lora_rank_64_bf16.safetensors HF Comfy-Org/MOSS-Image 307 MB
moss_image_color_depth_control.safetensors (optional) loras/moss_image_color_depth_control.safetensors HF Comfy-Org/MOSS-Image 317 MB

moss-realism

File Destination Source Size
realismByStableYogiMoss_v15Int8.safetensors diffusion_models/realismByStableYogiMoss_v15Int8.safetensors Civitai version 3434061 (token) 5.3 GB
moodyMossMix_v41Int8.safetensors diffusion_models/moodyMossMix_v41Int8.safetensors Civitai version 3449754 (token) 5.0 GB

moss-turbo

File Destination Source Size
lodestones-8-steps-dmd2-lora.safetensors loras/lodestones-8-steps-dmd2-lora.safetensors Civitai version 3420002 (token) 1.2 GB
moss_image_sft_4step_v2.0.safetensors loras/moss_image_sft_4step_v2.0.safetensors Civitai version 3414934 (token) 1.2 GB

ace-step

File Destination Source Size
acestep_v1.5_xl_turbo_bf16.safetensors diffusion_models/acestep_v1.5_xl_turbo_bf16.safetensors HF Comfy-Org/ACE-Step-1.5 10.0 GB
qwen_0.6b_ace15.safetensors text_encoders/qwen_0.6b_ace15.safetensors HF Comfy-Org/ACE-Step-1.5 1.2 GB
qwen_4b_ace15.safetensors text_encoders/qwen_4b_ace15.safetensors HF Comfy-Org/ACE-Step-1.5 8.4 GB
ace_1.5_vae.safetensors vae/ace_1.5_vae.safetensors HF Comfy-Org/ACE-Step-1.5 348 MB
acestep_v1.5_xl_sft_bf16.safetensors (optional) diffusion_models/acestep_v1.5_xl_sft_bf16.safetensors HF Comfy-Org/ACE-Step-1.5 10.0 GB
<!-- gmanski-packs:docs:end -->

Volume sizing

Volume Fits
100 GB node packs + Ollama + zimage + upscale + music
200 GB the above plus flux1, sdxl, captioning
250 GB+ qwen-edit (59 GB, bf16) or all

First boot

  1. Custom nodes install (2-5 minutes), ComfyUI starts, the detector models (~1 GB) download.
  2. Requested packs download in the background at 30-80 MB/s: about 10 minutes for zimage, 20 for music, 30 for qwen-edit. A workflow shows a "missing models" dialog until its pack has landed; do not use the dialog's Download button (it writes to the container disk).
  3. Later boots take about a minute; everything lives on the volume.
  4. Your JupyterLab link and FileBrowser password are in the box at the end of the pod log — see Logging in above. Port 8888 showing "Initializing…" in RunPod is normal.

Where the workflows are

Video generation tips

Preview cheap, then upscale. Rendering a MiniMax H3 clip directly at your final target resolution costs more time and, per community benchmarking, doesn't even get you a sharper version of the clip you wanted — same seed at a bigger size is a different clip (different camera, different background). Instead: generate 2-3 candidates at a reduced size/duration with a new seed each time (Gmanski Flow Simple/Advanced), pick the one you like, then run it through Video Upscale ESRGAN or Video Upscale LTX to sharpen it up. Both work on any video, not just this pod's own output.

No MiniMax-H3-specific latent upscaler exists. Checked directly against ComfyUI core's source (comfy_extras/nodes_minimax_h3.py) — there is no H3-native latent-space upscale node, official or third-party. latent_upscale_models (used by Video Upscale LTX) is LTX-2.5-specific, a different latent space entirely. For H3 output, Video Upscale ESRGAN's pixel-space route is the real answer, not a fallback.

Recommended GPUs

Use GPU
Z-Image, SDXL, upscaling, music 24-32 GB (RTX 4090 / 5090 / A5000)
Qwen Edit 2511 dataset generator (bf16) 48 GB+ (A40 / A6000 / L40 / RTX 6000)
On-pod LoRA training (training pack) 48-80 GB

Support

Workflows and nodes ship with this pod (Template Library and the Workflows sidebar, both under Gmanski/). Read !1 Gmanski Welcome and !3 Troubleshooting in the Workflows sidebar. Logs: /workspace/logs/.

Guides, more workflows and help: https://www.gmanski.com/runpod · email casting@gmanski.com.