Reference capture for MiniMax H3: one file per character, object or place that replays H3's reference conditioning exactly.
Capture a character's reference photos once, then replay them in any shot: no photo loaders, no resizing, no VAE encode, and conditioning bit-identical to the native reference node.
Features • Examples • Installation • Quick Start • Visual Guide • Nodes • Example Workflows • Models • Documentation • Troubleshooting • Contributing
- 🎉 First public release: RefCap capture and exact replay, view sets with clips and audio, multi-RefCap packs with a thumbnail picker, and 16 example workflows in ComfyUI's template browser.
One Photo, Any Shot: Consistent MiniMax H3 Characters. (Click to watch on YouTube)
|
H3 RefCap Replay hands the model exactly what MiniMax H3 Reference to Video makes from the original photos: in the tests its clips were pixel-identical to native references. Only the photo loading, resizing and VAE encode are skipped. |
Photos are stored as PNGs and resized again at replay, bit for bit: a 1024 px photo costs about 0.7 MB instead of 12 MB. The text encoder's cached rows are optional, and only Apply (cached) needs them. |
|
Many views of one subject, each with the angle it was seen from. Replay View Set picks the views nearest the shot's angles, and a view set can also carry up to 3 clips with soundtracks and 3 audio references. |
Two characters and a location in one shot: chain H3 RefCap Select nodes, each with a thumbnail picker, into Replay Pack. Bake Pack saves a cast you keep using as one file. |
|
Orbit Views picks a turnaround from two H3 orbit renders by the head yaw insightface measures, or from two camera pans round a place by their sideways motion. Workflow 8 turns one photo into RefCaps in a single queue. |
H3 RefCap 1-16 go from capture to motion, camera and location references, and workflow 16 runs on ready-made fantasy RefCaps. Each carries a note with its steps and what the tests measured. |
|
Replay: the elf ranger from her 5-view RefCap, walking a forest path she was never photographed in. |
Replay Pack: three views each of the elf and the dwarf, plus three of the crystal cavern, in one shot. |
Rendered at 608×352 with a community MiniMax H3 turbo finetune (int8, 8 steps) on a 12 GB GPU. The example workflows default to the official model.
- ComfyUI with the core MiniMax H3 nodes (tested on 0.41.0)
- The MiniMax H3 models (Models)
- insightface (optional, for Orbit Views'
facesmeasure)
Search for ComfyUI-RefCap in ComfyUI Manager and click Install, or run comfy node install comfyui-refcap.
Linux / macOS
cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/AEmotionStudio/ComfyUI-RefCap.gitWindows (PowerShell)
cd C:\path\to\ComfyUI\custom_nodes
git clone https://github.com/AEmotionStudio/ComfyUI-RefCap.gitNote: The nodes need nothing beyond ComfyUI, so there is nothing to
pip install.
Restart ComfyUI after installation.
- Open H3 RefCap 1 - Capture from ComfyUI's template browser (listed under ComfyUI-RefCap).
- Load up to 9 photos of one person, and connect them in the same order to both the reference node and H3 RefCap Capture. Leave the reference node's prompt empty.
- Set width and height to the size you'll generate at, give the RefCap a
name, and queue. It's saved asComfyUI/models/refcaps/<name>.safetensors. - Open H3 RefCap 2 - Replay (exact), refresh the browser, pick your RefCap in
refcap, and write the prompt with the photos as<Picture 1>...<Picture N>. - Queue. The latent comes from Empty MiniMax H3 AV Latent, as with native references.
💡 Tip: For the most identity, capture from photos of about 1024 px with
ref_image_size=max: +0.06 face similarity in the tests, at about 2.2x the render time.
H3 RefCap 1 - Capture: five photos go to the reference node and to H3 RefCap Capture in the same order, the reference node's prompt stays empty, and one queue writes the RefCap. Capture summary shows what it stored.
H3 RefCap 2 - Replay (exact): pick the RefCap, write the prompt with the photos as <Picture 1> ... <Picture 5>, and queue. The model gets exactly what the reference node would make from the original photos.
H3 RefCap 8 - Photo to RefCap: H3 renders two orbits round the character, one past each side; Orbit Views picks the frame nearest each angle by measured head yaw; and Build View Set saves a 24-view turnaround and a 5-view front RefCap. A full-body section does the same from a full-body photo, and a test shot uses both.
The 24 views it picked, and the test shot
H3 RefCap 7 - Two characters: one H3 RefCap Select per RefCap. Pick views... opens a thumbnail picker of everything in models/refcaps, with each view's angle, a search box, and a running count against the 9-image cap.
The Select nodes chain into Replay Pack, and its label guide says which <Picture i> is whom, so the prompt can name them.
All nodes are in the advanced/refcap category.
| Node | What It Does |
|---|---|
| H3 RefCap Capture | Captures H3 image references (latents, the pixels the encoder saw, optionally its rows) into one file |
| H3 RefCap Replay | RefCap + prompt -> conditioning identical to the native node fed the original photos |
| H3 RefCap Apply (cached) | Applies a RefCap's cached rows and latents to an already encoded prompt |
| H3 RefCap Build View Set | A RefCap of many views (with yaws), up to 3 clips with soundtracks and 3 audios |
| H3 RefCap Replay View Set | Replays chosen views, clips and audio of a view set as native references |
| H3 RefCap Orbit Views | Picks a turnaround from two H3 orbit renders or camera pans: the frame nearest each angle |
| H3 RefCap Select | Picks views, clips and audios from one RefCap into a pack chain ("Pick views..." button) |
| H3 RefCap Replay Pack | Replays a pack chain from several RefCaps as one conditioning |
| H3 RefCap Bake Pack | Saves a pack chain as one view set |
| H3 Save Conditioning (debug) | Dumps a conditioning to output/ |
RefCaps live in ComfyUI/models/refcaps/ (.safetensors, metadata key h3_refcap).
All 16 are in example_workflows/, and ComfyUI's template browser lists them under ComfyUI-RefCap. Their file inputs name the test files; load your own photos, clips and RefCaps in their place. To try RefCap without capturing your own first, download the fantasy cast from Hugging Face into ComfyUI/models/refcaps/ and run workflow 16.
H3 RefCap 1-16: RefCaps
| Workflow | What It Does |
|---|---|
| 1 - Capture | Captures reference photos into a RefCap |
| 2 - Replay (exact) | Generates with a RefCap, identical to native references |
| 3 - Apply (cached) | Generates with a RefCap's cached rows, without the photos going through the text encoder |
| 4 - AB native vs RefCap | Checks a RefCap against the photos it was captured from, on your character |
| 5 - View set | Builds a view set and replays the views nearest the shot's angles |
| 6 - Clips and audio | Up to 3 clips with soundtracks and 3 audio references in one view set |
| 7 - Two characters | Two characters and a place in one shot, from a pack of RefCaps |
| 8 - Photo to RefCap | One photo to face, 24-view and full-body RefCaps through H3 orbit renders, in one queue |
| 9 - Location | A location view set from one photo of a place, or from a description |
| 10 - Audio to RefCap | A voice RefCap from one sample |
| 11 - Video to RefCap | Stores one clip in a RefCap |
| 12 - Motion | A character performs a move stored as a clip RefCap |
| 13 - Camera move | A clip RefCap as the camera-movement reference |
| 14 - Edit a clip | Fixes one thing in a finished clip and keeps the rest of the take |
| 15 - Location walkthrough | A location RefCap from a video walking through a place |
| 16 - Fantasy cast | The elf, the dwarf, the crystal cavern and both voices from downloadable RefCaps |
The workflows use the official MiniMax H3 files from Comfy-Org/MiniMax-H3, with the 4-step Lightning LoRA as in ComfyUI's reference-to-video template. For the base model, bypass the Lightning LoRA node and use 20 steps.
| Model | Size | Stored In | Download |
|---|---|---|---|
| MiniMax H3 ref2va (int8) | 19.53 GB | ComfyUI/models/diffusion_models/ |
minimax_h3_ref2va_pruned_int8_convrot.safetensors |
| Qwen3-VL 32B text encoder | 14.61 GB | ComfyUI/models/text_encoders/ |
qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors |
| Video VAE | 2.62 GB | ComfyUI/models/vae/ |
minimax_h3_video_vae_int8_convrot.safetensors |
| Audio VAE | 577.2 MB | ComfyUI/models/vae/ |
minimax_h3_audio_vae_fp32.safetensors |
| Lightning LoRA (4 steps) | 1.82 GB | ComfyUI/models/loras/ |
minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors |
| insightface buffalo_l (optional) | ~300 MB | ComfyUI/models/insightface/ |
Downloaded by insightface on first use. |
| Doc | What's In It |
|---|---|
| REFCAP_README.md | The full write-up: how a reference reaches the model, the file format, every node, and the evaluation |
| REFCAP_VS_REFMOD.md | RefCap against ComfyUI-MiniMaxH3Mod's RefMods |
| REFCAP_AV_EXPLAINER.md | Clips and audio in view sets |
"encode the references with an empty prompt"
H3 RefCap Capture stores only the references. Clear the prompt on the MiniMax H3 Reference to Video node that feeds it.
"N images connected for M references"
Connect the same photos, in the same order, to the reference node and to Capture, or connect none to Capture and turn on store_rows (an Apply-only RefCap).
"has no stored pixels"
The RefCap was captured without its photos, so only H3 RefCap Apply (cached) can use it. Capture it again with the photos connected to H3 RefCap Capture.
"has no cached rows"
Apply (cached) needs the text encoder's rows. Capture again with store_rows on, or use H3 RefCap Replay, which needs no rows. View sets never store them.
"N views selected; MiniMaxH3ReferenceToVideo takes at most 9"
A render takes up to 9 images, 3 clips and 3 audio references, as the native reference node does. Pick fewer views in Replay View Set or with the Select picker, whose counter warns when a selection goes over the cap.
The RefCap doesn't show up in the dropdown
RefCaps are read from ComfyUI/models/refcaps/. Refresh the browser after capturing one.
Contributions are welcome! Whether it's bug reports, new workflows, or improvements — your help is appreciated.
- Fork the Project
- Create your Feature Branch (
git checkout -b feature/AmazingFeature) - Commit your Changes (
git commit -m 'Add some AmazingFeature') - Push to the Branch (
git push origin feature/AmazingFeature) - Open a Pull Request
This project is licensed under the GPL-3.0 License — see the LICENSE file for details.
Developed by Æmotion Studio
