Commit History
-
Sync the bundled nodes, workflows and docs with GitHub master (099a051)
Verified signatureThe mirrored copy had drifted to roughly the 5f40849 era, before the all-in-one work existed. That left the all-in-one checkpoint in this repo unloadable by the code shipped next to it -- Krea2SVDQuantCheckpointLoader was not in it -- and kept the dynamic-VRAM pin that made iteration time jump to 30-100 s past ~12 GB. Updates 12 files and adds 15, including build_all_in_one.py and the 12 workflow JSONs that were missing.06a0bf4
-
Sync docs and code: gallery, troubleshooting, cleanup pass
Verified signatured50078f
-
Sync nodes and scripts with the GitHub repo
Verified signatureBrings the code side up to date with three GitHub commits. Fixes a silent failure that mattered on the noLowRank checkpoint published here: the LoRA node detected quantized layers by looking for the low-rank branch, so on a branchless checkpoint it matched nothing, fell through to ComfyUI's normal LoRA path, and applied no LoRA at all to the 224 block linears without reporting an error. Adds the Quantize and Env Check nodes (five nodes total, now grouped under a Krea2/SVDQuant category), UI-format workflows that carry layout and a READ ME FIRST note alongside the API-format ones, tools/pixel_metrics.py for LPIPS/PSNR/SSIM against a BF16 reference, and a --rank-alloc option. Also makes refine_iters default to 100: measured over 10 prompts, LPIPS falls monotonically with rank only when refinement is on, and is flat without it.032e1ee -
Fix lowvram OOM, add backend diagnostics, support the base model
Verified signatureSyncs the node code with github.com/alperktt/Krea-2-SVDQuant-ComfyUI. - Low-rank factors are now visible to ComfyUI VRAM accounting and offload correctly, fixing OOM on small cards. - New diagnostics node/script: reports which comfy_kitchen backend actually runs. ComfyUI disables the CUDA backend on torch built against CUDA < 13, which silently makes these checkpoints slower than fp8. - LoRA branches are per-patcher object patches, so two LoRA nodes off one loader no longer contaminate each other. - quantize_krea2.py writes metadata, reports honest tensor counts, and fails loudly when no layer matched. - Base (non-turbo) model support: --variant flag and a base workflow.45bcac4
-
Upload workflows/krea2_svdquant_w4a4_t2i.json with huggingface_hub
Verified signaturebe1d214 -
Delete files svdquant_loader.py extract_base_parts.py fast_kernel.py workflows/krea2_svdquant_t2i.json with huggingface_hub
Verified signaturea964e27 -
Upload folder using huggingface_hub
Verified signature54e39db