← sources

Did a quick anime-to-realism test in MiniMax H3 with just a video reference + lora

source: reddit (reddit:StableDiffusion)published 2026-10-13 11:18ingested 2026-10-14 02:40

retained text

The post at the bottom inspired me to try a simple anime-to-realism test with MiniMax H3, so I figured I’d share a few results.

I didn’t generate any keyframes or stills. I just fed the original video directly into MiniMax H3 as the video reference, applied an anime to realism lora, and basically told it to convert it to live action.

The output quality was honestly better than expected. Not saying it’s perfect or amazing or anything, but for such a simple setup (ComfyUI official Minimax H3 Ref2V workflow + lora) I was pretty impressed.

I didn’t use any character reference sheets, so character consistency obviously isn’t guaranteed. It mostly seems to depend on how the model interprets the character in each scene.

For my next test, I might try adding a character reference sheet as well and see if that helps keep the character more consistent across scenes.

If anyone has any questions, happy to answer. Other output samples are in the comments.

Inspired by this post:  
[https://www.reddit.com/r/StableDiffusion/comments/1x1kvmg/recreated\_omnitrix\_calibration\_scene\_ben\_10\_mini/](https://www.reddit.com/r/StableDiffusion/comments/1x1kvmg/recreated_omnitrix_calibration_scene_ben_10_mini/)

LoRA used:  
[https://huggingface.co/LiseTY/Minimax-H3-ref2v\_Anime\_2\_Realism](https://huggingface.co/LiseTY/Minimax-H3-ref2v_Anime_2_Realism)

Here's the exact prompt I used:  
"Convert <Video 1>, the sole video reference, into natural photorealistic live-action footage. Follow the original shot frame by frame: preserve subject count, identities, camera, framing, positions, expressions, props, shot timing, and movement. Replace only illustrated surfaces with photographic skin, natural hair strands, real fabric and plausible physical lighting. Do not invent cuts, actions, subjects, or on-screen text. The result should appear filmed by a real camera, not cel-shaded or plastic CGI."

**Settings**:

**Parameter**	|**Value used**  
Checkpoint	|MiniMax\_H3\_Ref2VA\_pruned\_int8\_convrot.safetensors  
Workflow	|ComfyUI Offcial Minimax H3 Ref2V workflow + lora  
LoRA	|Minimax\_H3\_Anime2Realsim.safetensors  
LoRA strength	|1.0  
Steps	|20  
Sampler	|res\_multistep  
Scheduler	|simple  
Denoise	|1.0  
Video shift	|12.0  
Audio shift	|3.0  
Resolution during generation	|1024 × 576  
Frame rate	|24 fps

document fetched 2026-10-14 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1x4txte/did_a_quick_animetorealism_test_in_minimax_h3