lightx2v

/

Minimax-h3-Turbo

like 754

yes i only use R2v

Is there anything particularly great about RF? FL seemed to do well with image referencing, too.

It depends on what you're trying to achieve.
But you have way more control with Ref2v.

H3 Minimax is the first open-source model to allow real native ref2v the same way top models like Seedance and Kling do it and it changes everything.
This is a huge step forward for open-source AI.

So, what are the specific differences? Does it maintain multiple references better? Does it better capture the characteristics of the references? Does it handle video editing or sound references better? Probably, right? Then what are the downsides compared to FL?

I haven't tried video or sound yet, and I've only used a few subjects with a few images, but RF and FL felt pretty similar. Actually, I haven't tested it extensively. Of course, since I used a Turbo LoRA, it's a disadvantageous condition for RF. Anyway, right now FL seems better in terms of image quality—can an RF LoRA solve this?

since I used a Turbo LoRA, it's a disadvantageous condition for RF. Anyway, right now FL seems better in terms of image quality—can an RF LoRA solve this?

I had pretty good results with the Ref workflow on these specific settings :

Steps : 8
Sampler : dpmpp_sde
Scheduler : Beta

Sampler : dpmpp_sde

Great tip! I'm currently using a low-res (4) - latent upscale - high-res (2) steps setup with the Euler sampler. In the low-res phase, there wasn't much difference, but at high-res with low sigma, the effect of dpmpp_sde_gpu is really good.

You might not see it well, but the moiré artifacts are gone.
Left:dpmpp_sde_gpu / Right:Euler

Wait a second, there seems to be an error with ComfyUI's step time display. It shows the same amount of time taken, but it actually takes 3 steps worth of time. The result is also similar to Euler with 3 steps...
So if you use 8 steps of dpmpp_sde, it takes about as much time as 15 steps, so it must be good.

nice job thanks to the team. yeah the best if we could have a V1.0 4 steps and 8 steps Ref2va !!!