Open jordanjay29 opened 1 day ago
hr_forge_additional_modules
sounds like a good idea
You won't get good results with latent upscale between architectures, because the latents are not compatible. You can hiresfix any->any with a suitable upscale method, like Lanczos. I've tested 1.5 <-> XL (both ways) and 1.5 -> Flux. There is an issue if there is incompatibility between selected VAE/TE and the two models - i.e. trying to use a Flux model that doesn't include text encoders for the highres pass; because the necessary additional models are not included (unless also included in the first pass). Even 1.5 -> 1.5 can fail here, if the second pass model doesn't include a VAE.
I'm looking at adding highres additional modules. Currently working, but still some issues around infotext loading.
I expected that the output would be trickier to get right, but even having the ability to do so in Forge could at least make for some good experiments. Sometimes even bad output is good output, if you know what I mean.
A hires additional modules field would resolve it well!
Possible suggestion related to this, it may be nice if it could render the batch as one then hiresfix them as another. So render all 5 as SD 1.5 and then do them all as a hires batch with xl, to cut down on back and forth model loading for those of us with less resources.
I will say i'm happy to see this feature return
The recent hiresfix #2100 will only work across same architecture models (e.g. SD1.5 to 1.5, XL/Pony to XL/Pony, Flux to Flux) due to the change in requirements for VAE and possibly text encoders.
Could we have an additional field to select the TE/VAE like on top?