Commit Graph

Select branches

Hide Pull Requests

dev

main

v0.4.0

Mono Color

• e9e8b52d6f fix: use cached base model without re-verifying against HF (Xet 403) dev CthomasdesignandClaude Opus 4.8 2026-07-13 19:27:20 -04:00
• 8881d961c0 fix: model loader honors extra_model_paths (shared model stores) CthomasdesignandClaude Opus 4.8 2026-07-10 17:36:44 -04:00
• 2a7bb53800 tools: discover weights in the ComfyUI-Shared model store CthomasdesignandClaude Opus 4.8 2026-07-10 17:30:46 -04:00
• c9f8f0ace1 docs: README overhaul (architecture, editing guide, env vars, tools) CthomasdesignandClaude Opus 4.8 2026-07-10 17:16:08 -04:00
• 4d8db6bdf8 chore: declare macOS-only via classifiers; v0.7.1 CthomasdesignandClaude Fable 5 2026-07-09 13:59:11 -04:00
• 5fcf343a4d refactor(tools): machine-portable test scripts, shared asset helpers CthomasdesignandClaude Fable 5 2026-07-09 13:59:11 -04:00
• 1fc866c144 refactor(nodes): shared _run helper + widget factories, DESCRIPTION + tooltips CthomasdesignandClaude Fable 5 2026-07-09 13:59:11 -04:00
• b4ebb360e3 refactor(engine): tidy grounded additions — no behavior change CthomasdesignandClaude Fable 5 2026-07-09 13:59:11 -04:00
• 3ec843606c feat: two-reference grounded edits + preprocess robustness fix CthomasdesignandClaude Opus 4.8 2026-07-08 20:37:02 -04:00
• af579ffb6a feat: image-grounded edit conditioning via Qwen3-VL vision tower (v0.7.0) CthomasdesignandClaude Opus 4.8 2026-07-08 19:58:43 -04:00
• 1db2179abe feat(wip): vendor Qwen3-VL vision tower (MLX) + torch-free preprocessing CthomasdesignandClaude Opus 4.8 2026-07-08 19:31:17 -04:00
• 984aabc0d1 feat: Krea2Edit node — in-context image editing on Turbo/MLX (v0.6.0) CthomasdesignandClaude Opus 4.8 2026-07-08 18:52:53 -04:00
• 0f46e19cbc feat: Krea2Img2Img node — image-to-image on Turbo/MLX (v0.5.0) CthomasdesignandClaude Opus 4.8 2026-07-08 16:47:06 -04:00
• 98cb8b5380 docs: v0.4.0 — performance section, NOTICE amendment, version bump main v0.4.0 CthomasdesignandClaude Fable 5 2026-07-08 06:27:33 -04:00
• 151033af99 engine: prompt-embedding cache + encode-once for batches CthomasdesignandClaude Fable 5 2026-07-07 23:12:17 -04:00
• bfc60c11c5 engine: dynamic-length text encoding (541 -> ~prompt-length tokens) CthomasdesignandClaude Fable 5 2026-07-07 22:28:15 -04:00
• bf7b9cb7cf engine: trim padded text tokens before the DiT (~30% fewer tokens at 512²) CthomasdesignandClaude Fable 5 2026-07-07 22:08:01 -04:00
• 98ebf5b0bf engine: hoist step-invariant conditioning out of the denoise loop CthomasdesignandClaude Fable 5 2026-07-07 21:49:00 -04:00
• 80d0bf8f71 engine: use native GQA in fused attention instead of repeating K/V heads CthomasdesignandClaude Fable 5 2026-07-07 21:41:50 -04:00
• 59deb1e3e5 tools: benchmark + output-parity harness for engine optimization work CthomasdesignandClaude Fable 5 2026-07-07 21:07:07 -04:00
• 2ff0827d08 Initial commit: self-contained Krea-2-Turbo MLX nodes for ComfyUI CthomasdesignandClaude Sonnet 5 2026-07-07 20:32:37 -04:00