ComfyUI custom nodes for image, video and audio MiniMax H3 reference packs.
Python
7
17 commits
updated Oct 10, 2026
Build reusable MiniMax H3 reference packs from images, video and audio, then mix and select sources for each generation.
Save encoded references, subject descriptions and retention instructions together – spend less time preparing references and more time creating.
Features
.safetensors RefMods, avoiding repeated VAE encoding.Version 0.1.2 — an early release of the ComfyUI nodes. Interfaces and compatibility are still evolving.
Use a recent ComfyUI version with native MiniMax H3 and subgraph support. From your ComfyUI folder, run:
git clone https://github.com/entert-ai/ComfyUI-H3-RefMods-Lab.git custom_nodes/ComfyUI-H3-RefMods-Lab
Restart ComfyUI and refresh its browser page. Find the nodes under MiniMax H3 / RefMods Lab. The dependencies in requirements.txt are included in a standard ComfyUI installation.
Creating RefMods needs the appropriate H3 VAE. Generating video also needs an H3 Ref2VA model, its Qwen3-VL text encoder, and the visual/audio VAEs. Select your installed variants in the workflows; weights and source files are not included.
To update, run git pull inside the node's repository folder, then restart
ComfyUI and refresh the browser.
Drag either JSON into ComfyUI. The screenshots open at full resolution when clicked.
Open 01_create_refmod_Alice.json.
Alice. Give each source a description,
retention strategy and encoding resolution.Prepare Outfit1 with the same workflow: change Subject Name to Outfit1,
replace the images with outfit references, and replace Alice's descriptions and
retention details with clothing details. For example, describe its front view
and the cut, colour and fabric to preserve with fully_preserved. Disconnect
the unused Create branch from Combine if using one image, then queue to save.
The first saves are Alice_00001.safetensors and Outfit1_00001.safetensors.
Use the actual saved filenames if your counters are already higher. If Alice's
source clothing differs from the target outfit, choose partially_preserved
for that Alice source and specify that her identity stays while clothing may change.
Open 02_generate_with_refmod_Alice.json.
Include saved reference instructions is enabled. Text Encode adds the saved
subject definitions and retention analysis; Preview Any shows the exact final prompt.
With Alice first and Outfit second, both active, <Subject 1> is Alice and
<Subject 2> is Outfit1. Check that mapping whenever you change the references.
Resolution, length, seed and steps are exposed on the sampling subgraph. The example uses Euler/simple at 24 fps and selects Comfy Kitchen attention through the native Model Attention Backend node. If unavailable, ComfyUI falls back to PyTorch attention. Model choices remain editable.
| Node | Use |
|---|---|
| Create H3 RefMod | Encode images as individual reference sources. |
| Create H3 Video RefMod | Encode a video, optionally with its soundtrack. |
| Create H3 Audio RefMod | Encode an audio reference. |
| Combine H3 RefMods | Merge up to 100 packs with a strength slider per input. |
| Select H3 RefMod Sources | Include or exclude sources without re-encoding. |
| H3 RefMod Text Encode | Build the prompt, present references to Qwen and attach saved latents. |
| Set H3 RefMod Instructions | Edit subject descriptions and retention metadata without re-encoding. |
| Save / Load H3 RefMod | Store and reuse numbered .safetensors packs. |
| Inspect H3 RefMod | Preview sources, active reference labels and token usage. |
| Apply H3 RefMod (Latents Only) | Attach latents to existing conditioning without presenting media to Qwen. |
his, her or its, such as
her portrait. Descriptions are used as written.include_saved_instructions
in Text Encode and write the target video's action, shots, style and sound.
Remove manual [Subject Definitions] and [Retention Analysis] sections
when enabled. Use final_prompt to inspect the result.| Reference | Retention strategies |
|---|---|
| Image / video | fully_preserved, partially_preserved, attribute_transfer, weak_reference |
| Audio | fully_copy, partially_copy, reference, weak_reference |
unspecified leaves the strategy unset. Retention instructions guide the prompt;
they do not change latent strength or guarantee the outcome. Paired video/audio
has separate visual and soundtrack retention controls.
Text Encode already attaches the references. Do not also Apply the same pack.
Packs are saved to ComfyUI/models/h3_refmods_lab. Each filename prefix has a
persistent counter; deleting old files does not restart it. Back up the hidden
.h3-refmods-counter.sqlite3 file with your packs to preserve the counters.
Embed workflow is on by default. Drag a saved RefMod into ComfyUI to reopen its creation graph. Original source files are needed to rerun that graph, but loading the saved pack for generation does not require them. Embedded workflows include prompts and filenames; disable embedding before sharing if needed.
max_edge limits visual
resolution and max_tokens checks the reference budget.See technical notes for encoding details, API inputs, older-workflow migration and developer verification.
MIT — see LICENSE. Model weights and third-party dependencies retain their own licenses.
ComfyUI custom nodes for image, video and audio MiniMax H3 reference packs.
Python
7
17 commits
updated Oct 10, 2026
Build reusable MiniMax H3 reference packs from images, video and audio, then mix and select sources for each generation.
Save encoded references, subject descriptions and retention instructions together – spend less time preparing references and more time creating.
Features
.safetensors RefMods, avoiding repeated VAE encoding.Version 0.1.2 — an early release of the ComfyUI nodes. Interfaces and compatibility are still evolving.
Use a recent ComfyUI version with native MiniMax H3 and subgraph support. From your ComfyUI folder, run:
git clone https://github.com/entert-ai/ComfyUI-H3-RefMods-Lab.git custom_nodes/ComfyUI-H3-RefMods-Lab
Restart ComfyUI and refresh its browser page. Find the nodes under MiniMax H3 / RefMods Lab. The dependencies in requirements.txt are included in a standard ComfyUI installation.
Creating RefMods needs the appropriate H3 VAE. Generating video also needs an H3 Ref2VA model, its Qwen3-VL text encoder, and the visual/audio VAEs. Select your installed variants in the workflows; weights and source files are not included.
To update, run git pull inside the node's repository folder, then restart
ComfyUI and refresh the browser.
Drag either JSON into ComfyUI. The screenshots open at full resolution when clicked.
Open 01_create_refmod_Alice.json.
Alice. Give each source a description,
retention strategy and encoding resolution.Prepare Outfit1 with the same workflow: change Subject Name to Outfit1,
replace the images with outfit references, and replace Alice's descriptions and
retention details with clothing details. For example, describe its front view
and the cut, colour and fabric to preserve with fully_preserved. Disconnect
the unused Create branch from Combine if using one image, then queue to save.
The first saves are Alice_00001.safetensors and Outfit1_00001.safetensors.
Use the actual saved filenames if your counters are already higher. If Alice's
source clothing differs from the target outfit, choose partially_preserved
for that Alice source and specify that her identity stays while clothing may change.
Open 02_generate_with_refmod_Alice.json.
Include saved reference instructions is enabled. Text Encode adds the saved
subject definitions and retention analysis; Preview Any shows the exact final prompt.
With Alice first and Outfit second, both active, <Subject 1> is Alice and
<Subject 2> is Outfit1. Check that mapping whenever you change the references.
Resolution, length, seed and steps are exposed on the sampling subgraph. The example uses Euler/simple at 24 fps and selects Comfy Kitchen attention through the native Model Attention Backend node. If unavailable, ComfyUI falls back to PyTorch attention. Model choices remain editable.
| Node | Use |
|---|---|
| Create H3 RefMod | Encode images as individual reference sources. |
| Create H3 Video RefMod | Encode a video, optionally with its soundtrack. |
| Create H3 Audio RefMod | Encode an audio reference. |
| Combine H3 RefMods | Merge up to 100 packs with a strength slider per input. |
| Select H3 RefMod Sources | Include or exclude sources without re-encoding. |
| H3 RefMod Text Encode | Build the prompt, present references to Qwen and attach saved latents. |
| Set H3 RefMod Instructions | Edit subject descriptions and retention metadata without re-encoding. |
| Save / Load H3 RefMod | Store and reuse numbered .safetensors packs. |
| Inspect H3 RefMod | Preview sources, active reference labels and token usage. |
| Apply H3 RefMod (Latents Only) | Attach latents to existing conditioning without presenting media to Qwen. |
his, her or its, such as
her portrait. Descriptions are used as written.include_saved_instructions
in Text Encode and write the target video's action, shots, style and sound.
Remove manual [Subject Definitions] and [Retention Analysis] sections
when enabled. Use final_prompt to inspect the result.| Reference | Retention strategies |
|---|---|
| Image / video | fully_preserved, partially_preserved, attribute_transfer, weak_reference |
| Audio | fully_copy, partially_copy, reference, weak_reference |
unspecified leaves the strategy unset. Retention instructions guide the prompt;
they do not change latent strength or guarantee the outcome. Paired video/audio
has separate visual and soundtrack retention controls.
Text Encode already attaches the references. Do not also Apply the same pack.
Packs are saved to ComfyUI/models/h3_refmods_lab. Each filename prefix has a
persistent counter; deleting old files does not restart it. Back up the hidden
.h3-refmods-counter.sqlite3 file with your packs to preserve the counters.
Embed workflow is on by default. Drag a saved RefMod into ComfyUI to reopen its creation graph. Original source files are needed to rerun that graph, but loading the saved pack for generation does not require them. Embedded workflows include prompts and filenames; disable embedding before sharing if needed.
max_edge limits visual
resolution and max_tokens checks the reference budget.See technical notes for encoding details, API inputs, older-workflow migration and developer verification.
MIT — see LICENSE. Model weights and third-party dependencies retain their own licenses.