NRDX

Meet Henge — an open-source video studio named after Henge no Jutsu, the transformation technique from Naruto. Keep the performance. Transform the character. 🥷

Built on ComfyUI + MiniMax H3, Henge turns a guide video and reference images into Character Swap, Motion Transfer, Head Swap, Multiple Character Swap and Sharpness/upscaling results.

The existing features use LoRAs I previously released. The new Motion Transfer introduces new LoRAs I trained, including TARA, my temporal-alignment technique for following the guide’s motion.

The built-in AI assistant helps inspect footage, mark and assign multiple characters, prepare prompts and launch generations faster. You can also use your own AI client through MCP.

Run locally or on RunPod Pods and Serverless. Comfy Cloud integration is being validated and requires our Henge custom nodes, now published in the Comfy Registry, plus the required models and dependencies.

Watch the demo below

Open source. MIT-licensed app.

github.com/alisson-anjos/…

#OpenSourceHiggsfield

I’ve developed TARA — Temporally Aligned Reference Attention, an experimental method for MiniMax H3.

Research, implementation and video examples:

github.com/alisson-anjos/…

Native video references can capture an action but drift out of sync with the source. TARA aligns the reference and output timelines inside attention, helping preserve performance and timing while using native reference conditioning, without Add Guide.

I’ve also developed optional precompiled kernels for Comfy Kitchen and sol-attn, with TARA’s temporal attention bias fused directly into the attention computation. This avoids the slower per-frame compatibility path. BFSNodes automatically downloads the matching binary on supported systems, compatibility and performance across GPUs are still being validated.

Early tests look promising. Interestingly, my guide-latent-trained LoRA currently seems to work better with TARA than the unfinished TARA-trained checkpoint. This isn’t a controlled comparison; I’m training a fresh LoRA from scratch to investigate further.

There’s a TARA node in BFSNodes for native ComfyUI workflows. If you’d like to try it, feedback, settings and short comparisons are very welcome, especially from different GPUs!

Sorry folks, I know you're tired of Character Swap LoRAs 😅 but here's an update to my v1.0.

I spent the last week training version after version to improve likeness, and I think I landed on a really nice one. I almost called it v2.0, but for that I'd want multiple characters swapped at once. I don't doubt it's possible, I just haven't tested it. So v1.1 it is.

It's also much smaller than the previous one (80 MB vs 620 MB): I extracted the result with SVD at a dynamic rank, keeping 95% of the energy (Frobenius norm) per layer. So every layer has its own rank: the lowest is 4 and the highest is 34 (median 7).

Hope it really beats v1.0. Now back to my OpenJutsu for the Comfy challenge 🔥

Links:

huggingface.co/Alissonerdx/Mi…

huggingface.co/Alissonerdx/Mi…

If this LoRA proves useful for your work in any way, please consider helping me continue training more LoRAs like this one ☕ buymeacoffee.com/nrdx

Below are some tests I ran at a low resolution (0.3 MP); as a result, you might notice some artifacts and the like, which is perfectly fine, and likely due to the very low resolution.

#ComfyUI #Minimax #H3 #aigenerated #CharacterSwap @ComfyUI @MiniMax_AI @huggingface

The Character Swap LoRA is out! Feel free to try it and share what you make.

It’s far from perfect, it’s just a LoRA, but it can work surprisingly well in some cases, like the examples below. I hope to work on an improved version soon together with @akatz_ai.

LoRA:
huggingface.co/Alissonerdx/Mi…

ComfyUI workflow: huggingface.co/Alissonerdx/Mi…

And there's an AI-generated video (video, voice, but it's my real voice, just in English hahaha) here hahaha to explain the Workflow planner node and some other things about LoRa: youtube.com/watch?v=D037QM…

V1 of my body/character swap LoRA for MiniMax H3 is coming soon!

Still a work in progress! I'm refining it with reinforcement learning + custom alignment metrics (pose, expression, identity, framing).

It's a slow process (9 min/step), but we're getting there! 😅