Try a simpler process where the canvas is scaled, rotated and cropped square.

Update submit button to say "Generate".

Correct typo in desired output width docs.

Correct text2img_guidance_scale param name.

Update preview tool.

Update the tool to render half-drops.

Make text2img guidance scale a float.

Restructure README config.

More config for HF.

Add missing HF config to README.

Add README to configure spaces.

Improve test tool. Make deployable to a hf space.

Log out params used.

Update handler to extract all params. Remove memory logging.

Fix origina image width calculation.

Increase img2img strength to 0.6 to attempt tiling improvement.

Adjust parameters to address render issues.

Update diffusers dep on macOS.

Checkin half-drop testing tool.

Remove forced upcast (no longer needed).

Try a fixed fp16 VAE.

Explicitly enable upcasting, to confirm it is on.

Try explicitly disabling upcasting on the vae.

Re-cuda pipes per execution.

Restore VAE slicing (not supported by AutoencoderKL, so not a candidate for root-cause).

Disable VAE slicing. Flush GPU memory.

Revert "Try disabling FP16."

Try disabling FP16.

Upgrade diffusers to 25 on HF endpoint.

Correct response format for HF inference endpoint.

Move pipe to cuda call to ensure types are correctly modified.

Attempt to cast bias to f16.

Try disabling VAE slicing to see if that removes the CUDA incompatible half type.

Use multiple GPUs (for Nvidia A10G support).

Use a single GPU (for Nvidia A100 support).

Support device in pipe initialisers.

Place pipelines on different GPUs.

Additional VRAM debugging.

Touch VRAM debugging statements.

Use CUDA 11.8 compatible version of pytorch. Separate install args for macOS.

Update GPU identifying logic.

Implement a HF handler.

Select between cuda or mps based on environment.

Tweak params and prompts for diamond-crop-img2img.

Add func to compute tile width from wallpaper width.

Remove redundant self call.

Preserve image mode to ensure colour images remain colour.

Make experiments consistent & use shared utils.

Extract and refactor diamond crop img2img procedure.

Working prototype, generating half-drop from a diamond-crop.