← sources

[ H3 ] ref2va looks dev

source: reddit (reddit:StableDiffusion)published 2026-10-22 16:30ingested 2026-10-23 02:40

retained text

saw this post: [https://www.reddit.com/r/StableDiffusion/comments/1xbobrr/original\_character\_lora\_for\_minimax\_h3\_learnings/](https://www.reddit.com/r/StableDiffusion/comments/1xbobrr/original_character_lora_for_minimax_h3_learnings/)

Hey guys,

SeeDance videos look much more realistic and softer. I'm really interested in pushing H3 in that direction, but ref2va gen always looks bad - dark, high contrast, and oversaturated.

I saw the post above, great info and results. However, I want to avoid doing any LoRA training. I think with the H3 FL2VA workflow you can do it much better as long as you have a good start image where you control the look. But I want to focus on making ref2va work.

In my last 3-minute duration test, I was using the fizgig node to make the skin look softer. It works, but the effect keeps accumulating in motion-context shots. That node does it in latent space, but I think it's more controllable and easier to do it in image space so the effect doesn't accumulate. So I made an image-space version with AI.

So it's a hack - if you bring up the dark values there's still no detail. I'll keep looking into it. I'm thinking maybe LTX SDR to HDR could help, since we get more details in bit depth after gen, That could be a way.

Of course, please share your idea on this, how you get better result with ref2va. Thanks!

If you guys find this more handy than doing it in post / compositing software, I'll share it.

2 nodes I am using "Image Detail Contrast" and "Orion4D Tone Mapping". The setting I have just over the top here.

Cheers.

\- the gen is using hybrid model loader with native fl2va ref2va, b5-49, WITHOUT using any turbo lora.

\- I've tried all the LoRAs, like Realistic Peoples and others, but I just can't get the result I want.

\- You can check the overlay text of some infomation.

\- I haven't test this with motio-context... hope it won't have issue.

TLDR: This is a typical image unsharp mask to get a softer skin look, plus some color grading. The point is not do grading in post, it is how we could get H3 ref2va workflow get better result in the first place.

document fetched 2026-10-23 00:34 · canonical https://reddit.com/r/StableDiffusion/comments/1xcnefw/h3_ref2va_looks_dev