Introduction

Swap the star of any clip, keep every gesture. That's the promise of the Character Swap LoRA for MiniMax H3, and its creator Alissonerdx (NRDX) just pushed v1.1 to Hugging Face on October 18. You feed it a guide video plus one reference image, and H3's ref2va model re-renders the main subject as your character while keeping the original scene, camera, pose, lip movement and facial performance.

v1.1 is all about likeness. After a week of training run after run, the new checkpoint keeps the face, hair, body and clothing closer to the reference than v1.0 did. It's also tiny: 80 MB versus 620 MB. It runs locally in ComfyUI today with two ready-made workflows, and the weights are free to use, including for commercial videos.

What's new in v1.1

  • Better identity hold. The generated character drifts less from the reference image across frames, while still following the guide's motion and framing.
  • About eight times smaller. v1.1 is an SVD extraction with a dynamic rank per layer. Each of its 208 layers keeps the smallest rank that holds 95% of its energy (Frobenius norm), so ranks range from 4 to 34 with a median of 7. v1.0 was a flat rank-64 file.
  • Strength 1.0 is the recommended setting.
  • v1.0 stays in the repo for side-by-side comparisons.

Alissonerdx says a true v2.0 would swap several characters at once, which hasn't been tested yet, so treat v1.1 as a one-subject tool.

The two ComfyUI workflows

Both workflows render in 8 steps with H3's DMD turbo LoRA, and both include an optional second pass with the LMS LoRA plus the H3 latent upscaler at 1280×704.

  • Shot Planner workflow (minimax_h3_character_swap_workflow_v1.1.json): built on ComfyUI-BFSNodes 1.113.0 or newer. It splits long videos into shots, renders each one, joins them, and outputs a side-by-side comparison.
  • Native nodes workflow (minimax_h3_character_swap_workflow_native_nodes_v1.1.json): core ComfyUI only. A Generate Text node looks at your reference image and writes the full prompt, then MiniMax H3 Reference to Video plus Add Guide render one clip.

Models the workflows load:

  • minimax_h3_ref2va_pruned_int8_convrot (H3 ref2va diffusion model)
  • qwen3vl_32b_minimax_h3_nvfp4_awq text encoder from Comfy-Org (int8 and bf16 versions also work)
  • minimax_h3_video_vae_fp16 and minimax_h3_audio_vae_fp32
  • minimax_h3_dmd_ref2va_8step_turbo_pruned (8-step turbo LoRA)
  • qwen3.5_9b_w4a8_gptq or any Qwen3-VL / Qwen3.5 model that Generate Text supports, to describe the reference
  • Optional: minimax_h3_lms_v1.0_r64 and h3_upscaler_lms_v0.1 for the second pass

Each workflow carries a note with download links for everything it loads.

Quick setup

1. Drop in the LoRA

Download the v1.1 file and place it in your loras folder:

huggingface-cli download Alissonerdx/Minimax-H3-ComfyUI loras/minimax_h3_character_swap_v1.1_ref2va_fro.safetensors --local-dir ComfyUI/models

That lands it in ComfyUI/models/loras/. Grab a v1.1 workflow from the repo's workflows/ folder the same way.

2. Wire the inputs

  • Send the original clip into MiniMaxH3AddGuide at frame_idx = 0. The LoRA was trained on guide-latent conditioning, so this is the recommended video path.
  • Send the character image through H3's native image-reference path as <Picture 1>.
  • Keep the output aspect ratio and frame count matched to the guide, using 17n + 5 frames.
  • If the guide has several people, say which one to replace, or the most prominent subject gets picked.
  • Start with a short test clip before you commit to a long range.

3. Use the v1.1 prompt template

Paste this on one line and replace only {details}:

body_swap: subject_definitions: <Subject 1> is the {details} shown in <Picture 1>, whose identity, face, head, hair or fur, skin, body shape, proportions and clothing define the character in the target video. The guide video provides the scene and the performance: its background, environment, lighting, camera, framing, body pose, hand and arm movement, head rotation, gaze direction, lip movement and facial expression, frame by frame; only the main subject of the guide video is replaced: its face, hair, body and clothing become <Subject 1>.  summary: [video generation + reference] Replace the guide subject with <Subject 1>, keeping the scene, camera and the complete motion and performance of the guide video.  detailed_description: [Shot 1] <Subject 1> takes the exact place of the guide subject and moves with the same timing, body pose, gestures, head motion, gaze, speech and facial expressions, while keeping the complete identity from <Picture 1>: facial structure, hair or fur, skin, body shape and clothing stay stable in every frame. The background, lighting, colors and camera of the guide video are preserved. Prevent identity drift, clothing changes, texture crawling and flicker.

{details} is a comma-separated, head-to-toe description of your reference character. Start it with the gender, with no article in front and no full stop at the end, so it reads naturally between "is the" and "shown in". For example: Male, average height, slender build, oval face, blue eyes, long white beard, gray cap, dark jacket.

In the native workflow, Generate Text fills this in for you. Set its max_length to about 768 so the end of the prompt doesn't get cut off. In the Shot Planner, the Describe refs button in the Prompts & refs tab writes the description.

4. Optional: keep the guide's audio

In the Shot Planner, Use video audio picks the output soundtrack. Audio conditioning also needs with_audio = true on BFS Shot H3 Conditioning.

Tips and limits

  • Pair it with Head Swap. The Head Swap LoRA in the same repo uses similar guide-latent conditioning, and combining the two may help lip-sync on some clips. The author flags this as still experimental.
  • Expect drift on hard shots. Small faces, occlusion, fast motion, tricky angles, lighting changes and long clips can all loosen likeness.
  • People first. It targets humans and human-like characters. Animal references are less tested.
  • The author's demos ran at 0.3 MP, so some artifacts in them come from the very low test resolution.

License in plain words

The Character Swap weights use the custom BFS Character Swap Community License 1.1, not an OSI open-source license. Personal, research and commercial use are allowed, and you can sell the videos and films you make. Organizations with more than US $1 million in annual revenue must credit the LoRA and link the model page. You can't resell the weights as a standalone product or offer them as a hosted API or SaaS service. MiniMax H3 and ComfyUI keep their own terms, and you need rights and consent for every face and clip you feed in. Please don't use it for impersonation.

Original Source

Conclusion

v1.1 is the kind of update creators love: the same one-reference recast trick, a tighter grip on who your character is, and a file small enough to stop hogging your drive. Load the native workflow, run a short clip, and watch your character step into the scene without missing a beat.

Want to play with base MiniMax H3 before setting up locally? Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&amp;utm_campaign=gen&amp;utm_content=minimax-h3-character-swap-lora-v1-1-comfyui

—Aurelia ♡