corrected requirements_versions.txt adjustments (otherwise wouldn't work if using README.md parameters)

[suphamster]

accelerate==0.26.0
diffusers==0.32.0

P.S. Btw processing speed is very slow at NV RTX 4070 and about 1/4 of its power is used.

[DenOfEquity]

Thanks for correction.

You could try editing sampler_invsr.py in the huggingface_hub directory. Lines 115-116:

        sd_pipe.to("cuda")
#        sd_pipe.enable_model_cpu_offload()

This improves GPU utilization from ~50% average to ~95% on my 1070 but then requires >8GB VRAM for medium size or larger images, otherwise will spill into shared memory and be much slower. It'll need a full restart of Forge to take effect after making the change.
Currently investigating modifications to the pipe - seems like it encodes the prompt for every tile.

[DenOfEquity]

I've reworked the model handling: now only encodes the prompt once, VAE encodes all tiles, then inferences all tiles, then VAE decodes all tiles - so reduced VRAM usage + much less model moving on GPUs with low VRAM. (original HF Space implementation, for each tile: encode prompt > VAE encode > inference > VAE decode; with potential CLIP/VAE/unet moving before each stage).
Also uses Forge's model movement, which should be better than diffusers model offloading, although this might not make any real difference after the main changes.

[suphamster]

Thanks, now it works much better/faster. Anyway I think this SR is a bit unpolished (not to get results as described here zsyOAOA/InvSR#5) and needs good quality input (if upscale real images for example as separate tool at spaces tab) and better faces handling but it's question to its developer and not you.