🎬 Upload your results! 💡 They give me new ideas on what to improve. The workflow is inside the mp4 — just drop it in the gallery 📥 Failures welcome too 🐛🔧
# v1.1 update
## Form fields (new version)
- **Version name:** v1.1 — 🖼️ Image to Video + ✨ Refine
- **Files:** H3_LongTake_workflows_v1.1.zip (4 workflows + README)
- **Gallery (this version):** 08_i2v.mp4 (image to video + refine from one picture, 15 s, source picture on the left)
- **Extra tags:** image to video, hires fix
## Version description (paste into the version's "About this version" box)
🚀 **v1.1 — Image to Video + Refine** (needs node pack **1.1.1+** → update from ComfyUI-Manager)
🖼️ **New workflow H3_LongTake_i2v — a long video from a single picture.** start_image is the first frame,
prompts holds **one block per 5-second clip**, separated by a --- line (block 1 = 0–5 s, block 2 = 5–10 s…,
the last one repeats). One clip in memory whatever the length; 🔊 the audio is generated by H3 per clip.
🧑🦰 **Identity references against drift** (ref2va model): the picture as <Picture 1> and a close-up of the face as
<Picture 2>. 📏 Measured on 30 s: without a reference the face becomes another person within 20 s (similarity
0.27), with the face reference it stays at 0.60.
✨ **New node Refine (second pass) — video "hires fix"** on any finished project (video to video or image to video):
0.5 MP → 1.0 MP with denoise 0.25, real skin and hair detail, and with the picture + face connected it puts the
picture's face back on the whole video even where the first pass lost it (📏 0.27 → 0.62). Output in <project>/hr,
assembled by the Stitch like any project. ⏱️ About 6 minutes per clip on 16 GB.
📝 **README note inside every workflow** (what it does, settings, steps) — and tidier boxes and titles.
📦 **Extra models for the i2v workflow:** minimax_h3_fl2va_pruned_fp8_scaled.safetensors +
minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy.safetensors (no references), or the ref2va model + Turbo ref2v of
the other workflows (with references).
📖 Full docs, measurements and prompting guide: https://github.com/mark9009/ComfyUI-H3-LongTake
## One line to add at the top of the main description (optional)
🆕 **v1.1:** 🖼️ Image to Video workflow (a long video from one picture, one prompt block per 5 s) and ✨ Refine
second pass at 1.0 MP — see the version notes.
Three ComfyUI workflows for MiniMax H3 that turn a source video of any length into a new video, clip by clip,
on a single 16 GB GPU. They need the custom node pack ComfyUI-H3-LongTake (MIT, no extra Python dependencies):
👉 https://github.com/mark9009/ComfyUI-H3-LongTake — install, models list, tested use cases with the exact prompts, prompting guide.
What the node does
- Slices the source video into 124-frame clips (5.2 s) that overlap by 5 frames, decodes only each slice with ffmpeg, and renders them one after the other.
- Every clip gets the previous clip's latent tail as an H3 keyframe, so the seam is continuous (measured 0.96 continuity) and the video follows the source with no drift over minutes of material.
- Resumable projects: stop ComfyUI, queue again, it continues from the first missing clip. Redo a single clip redo_one) or everything from clip N redo_from) with a new seed while the seams stay.
- seam_match = color removes the colour/exposure jump between clips (ΔE 5.8 → 0.9).
- The Stitch node concatenates the clips without re-encoding, puts the original audio back and embeds the workflow in the mp4 (drag the result into ComfyUI to reopen the project).
The three workflows
1. H3_LongTake_character_swap — a picture's character performs the whole video source_role = reference, Ref2VA). Identity from <Picture 1>, motion, timing, camera and room from the video. Works with photoreal and anime references; the video's room is kept even when the picture has its own background (say so in the prompt).
2. H3_LongTake_example — the same, with the structured H3 prompt (subject definitions / retention analysis) for motion + identity transfer.
3. H3_LongTake_style — style transfer and retexture with the StyleTransfer LoRA by NRDX (https://civarchive.com/models/2932297): the video becomes a guide latent source_role = guide), motion and framing are reproduced frame by frame (adherence 0.93), the style comes from text (Van Gogh, Ghibli, Simpsons, GTA…) or from a style swatch as <Picture 1>. retexture: changes the material/colour of one garment and nothing else.
Models (all from the official Comfy-Org MiniMax-H3 release unless noted)
- minimax_h3_ref2va_pruned_fp8_scaled.safetensors (diffusion_models)
- a MiniMax H3 Qwen3-VL-32B text encoder build (int8 or nvfp4; ~25 GB, kept in system RAM)
- minimax_h3_video_vae_fp16.safetensors, minimax_h3_audio_vae_fp32.safetensors (vae)
- minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors (loras) — 4 steps
- optional: minimax_h3_style_transfer_v1.0_r64.safetensors by NRDX (loras) — style workflow only
Quick start
1. Load a workflow, put your video in ComfyUI/input/ and pick it in source_file; connect your picture to ref_image_1.
2. dry_run = true → the node prints the slicing plan.
3. dry_run = false, max_clips = 2 → check identity/style and the seam in the preview.
4. max_clips = 0, mode = continue → the whole video. Unmute the Stitch node (Ctrl+M) when all clips are done.
Timing on a 16 GB GPU at 544×960: about 4:40 per 5-second clip with 4 Turbo steps (~1 minute of video per hour).
Prompting rules that came out of the tests
- Named styles work from text alone: 3–5 visual attributes + "Keep the original room, background, framing and motion exactly as in the video; do not add any new scenery." Never a style name and <Picture 1> in the same sentence.
- Style pictures without faces or scenery (crop a swatch): a face transfers the identity, a skyline replaces the room.
- Character swap: describe the character in words as well as <Picture 1>, and forbid the picture's background explicitly. A picture framed like the video keeps the motion in sync.
- retexture: names the exact garment ("the dress", not "the outfit").
Examples in the gallery (source | result): character swap from a photoreal picture, Van Gogh, Studio Ghibli, GTA from a swatch, The Simpsons on a 16:9 press conference, pop-art from a halftone swatch, gold-armour retexture. Source clips are free Pexels videos and a film excerpt; the character reference is an AI-generated picture.
Credits: motion-context idea by NikoDemon80 and the Banodoco MiniMax H3 thread; StyleTransfer LoRA by NRDX; MiniMax H3 core nodes by Comfy-Org. Node pack and workflows: mark9009.
Description
# Civitai — v1.3 update only
## Form fields (new version)
- Version name: v1.3 — ⚡ latent upscaler in Refine + 🧑🎨 character sheets
- Files: H3_LongTake_workflows_v1.3.zip (4 workflows + README)
- Gallery (this version): reuse 08_i2v.mp4 or add a new Refine clip if you have one
- Extra tags: latent upscale
## Version description (paste into the version's "About this version" box)
🚀 v1.3 — faster, sharper Refine and character sheets (needs node pack 1.3.0+ → update from ComfyUI-Manager)
⚡ Refine with the H3 latent upscaler. The second pass now enlarges the latent instead of decoding,
resizing and re-encoding the video upscale = latent_model, the default). 📏 Measured per clip, same
references and seed: 253 s instead of 375 s, face sharpness 222 vs 85, same identity, hands +60 %.
📦 Install the Comfyui_Minimax_h3_latent_Upscaler node pack and put
minimax_h3_latent_upscaler_3d_fp16.safetensors in models/latent_upscale_models. Without them the Refine
falls back to the old pixel upscale and tells you in its report — nothing breaks.
🎞️ Refine from edited chunks frames_from = mp4): refine clips you already touched up in pixels (a face
fix, a colour grade) instead of the raw latent.
🧑🎨 Character sheets (from 1.2.0, never announced here): character_sheet on Image to Video and Refine, a
sheet with several views of the character as an extra reference. 📏 Use it on top of the picture + face
references: alone it is worse, together it raises the identity floor (min 0.16 → 0.26 on a hard 15 s test).
🧠 Newer ComfyUI (comfy-aimdo dynamic VRAM) fixed: the second run in the same session no longer slows down
until it hangs — the nodes free VRAM before each heavy phase.
🛡️ Safer dry runs: with the Render in dry_run, Refine and Stitch stop with a message instead of
re-assembling an older project with the same name. project_name is greyed out while project_dir is linked.
🔧 Fixed in the i2v workflow: the Stitch's audio_file held an Italian value that failed validation once
you unmuted it. The Refine node now shows the new upscale / frames_from options and its README note
explains the latent upscaler.
📖 Full docs, measurements and prompting guide: https://github.com/mark9009/ComfyUI-H3-LongTake
## One line to add at the top of the main description (optional)
🆕 v1.3: ⚡ Refine with the H3 latent upscaler (a third faster, much sharper faces) and 🧑🎨 character
sheets — see the version notes.
FAQ
Comments (4)
H3LongTakeFaceRefine isn't in any public version of ComfyUI-H3-LongTake.
The code has no FaceRefine node. longtake_nodes.py defines only Render, ImageRender, Refine and Stitch. There's no object_image input either, and the repo's own H3_LongTake_i2v.json example doesn't use FaceRefine.
The workflow came from a modified, unreleased version. It labels its nodes as comfyui_fearnworksnodes 0.1.2 (a version that doesn't exist publicly) and adds FaceRefine, the logo input and a few extra settings. Someone extended LongTake privately, and only they have that version.
So the full workflow as designed can't be reproduced from public code. The FaceRefine building blocks (Face Track + Crop, Stitch Back and so on) are installed, so the step could in principle be rebuilt by hand
Thanks for the information. I noticed that two packages I use modify a node-loading phase, making everything look like part of their nodes; I’ll disable them and report the issue to the developers. As for my own node, that was my mistake: I develop in a different language on a separate branch, so I didn't pull in all the changes from the latest version—I'll update the public nodes shortly.
Try one of the latest videos.
Those transitions are still very, very rough