AlperKTS/Krea-2-SVDQuant-ComfyUI

Add model card metadata (enables download tracking)

Sync the bundled nodes, workflows and docs with GitHub master (56adad3)

Sync the bundled nodes, workflows and docs with GitHub master (9d1bb8b)

Sync the bundled nodes, workflows and docs with GitHub master (9907421)

Sync the bundled nodes, workflows and docs with GitHub master (97a0cf1)

Sync the bundled nodes, workflows and docs with GitHub master (ba4d62d)

Sync the bundled nodes, workflows and docs with GitHub master (0e6c8e6)

Sync the bundled nodes, workflows and docs with GitHub master (7d6a043)

Sync the bundled nodes, workflows and docs with GitHub master (099a051)

Add the rank-256 act-aware all-in-one checkpoint (DiT + 4-bit Qwen3-VL 4B text encoder + VAE, 12.77 GiB)

calibration: drop the old ambiguous name (renamed, see calibration/README.md)

calibration: drop the old ambiguous name (renamed, see calibration/README.md)

calibration: document which file is which, and which one not to reach for

calibration: name the LoRA capture after the adapter it was captured with

calibration: rename the turbo capture to say which model it describes

calibration: add the LoRA-loaded turbo capture (bloomgirls, lora2 arm)

Sync the node pack: compile-friendly kernel op, faster conversion, honest int8 numbers

fix: bake_adapter offset, plus a preflight and a discoverability hint

feat: adapters dial on the LoRA node, and bake_adapter.py

perf: measure the LoKr bypass path, and keep it staging per call

feat: support LoKr, LoHa and OFT LoRAs on quantized layers

docs: list the base rank-256 actaware checkpoint; bench gains a base mode

add base activation statistics (8 prompts, BF16 raw, 2.6M tokens/layer)

add base rank256 activation-aware checkpoint

fix: keep sage off attention shorter than one K/V block

fix: accept the transformer. LoRA key prefix

fix: keep sage-attention off float attention masks

Sync docs and code: gallery, troubleshooting, cleanup pass

Sync docs and code: gallery, troubleshooting, cleanup pass

Move checkpoints into checkpoints/ and checkpoints/legacy/

Add calibration/krea2_act_stats.safetensors

Add Krea2-Turbo-SVDQuant-W4A4-rank256-actaware.safetensors

Add activation-aware low-rank objective (--act-stats) + Test 4 benchmark

Measure LoRA fidelity properly: branch rank matters under a LoRA

Fix LoRA stacking double-applying non-quantized layers

Tidy: sync README, rename cmp2_* to compare_* (matches GitHub)

Add Krea2-Turbo-SVDQuant-W4A4-rank16-norefine.safetensors

Add Krea2-Turbo-SVDQuant-W4A4-rank128-norefine.safetensors

Add Krea2-Turbo-SVDQuant-W4A4-rank16-norefine.safetensors

Add Krea2-Turbo-SVDQuant-W4A4-rank256-norefine.safetensors

Add Krea2-Turbo-SVDQuant-W4A4-rank256-norefine.safetensors

Add Krea2-Turbo-SVDQuant-W4A4-rank256.safetensors

Add Krea2-Turbo-SVDQuant-W4A4-rank256.safetensors

Delete speedtest3.bin with huggingface_hub

speed test 3, will delete

Delete speedtest2.bin with huggingface_hub

speed test 2, will delete

Delete speedtest.bin with huggingface_hub

speed test, will delete

Upload BENCHMARKS.md with huggingface_hub