Commit History
-
Update app.py
Verified signature0d5f418
-
Fit ZeroGPU public duration limit
5309b29 -
Pin accelerated transformer to owned mirror
0d7ad9a
-
Validate image input before GPU inference
3c1179c -
Remove unnecessary Space token dependency
3d5646c -
Prevent low-resolution grid artifacts
ae63a91 -
Quantize model stack for ZeroGPU memory limits
5af9119 -
Allow cold four-step edits to finish on ZeroGPU
2137a28
-
Harden privacy and pin four-step model stack
e01757b
-
Remove undisclosed prompt logging that secretly saves user prompts and images (#2)
- Remove undisclosed prompt logging that secretly saves user prompts and images (927594b1263a9597116addc87d60ac4e9105b1be) Co-authored-by: sigma <signsur4739379373@users.noreply.huggingface.co>272b5bf
-
Make model loading resilient for duplicated Space
0aaa760
-
Pin kernels<0.15.1 to dodge required revision/version arg break
mooseca7f544
-
Remove temporary dimension-flow debug prints
moose +1: Claude Opus 4.7 (1M context)Reverts the debug instrumentation from 7a46cb6 now that the output aspect ratio fix is confirmed working. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>ad7b834 -
Derive output aspect ratio from first image, not last
moose +1: Claude Opus 4.7 (1M context)Multi-image edits were inheriting the second input's aspect ratio because output dimensions were calculated from image[-1]. Use image[0] so the subject/canvas (slot 1) drives output size. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>09cc589 -
Add temporary debug prints for dimension flow (will revert)
moose7a46cb6 -
Revert "Preserve input image aspect ratio when sliders are at default"
mooseThis reverts commit efe093f967e923c5779f38f59bf90fde2afeb737.593798f -
Preserve input image aspect ratio when sliders are at default
moose +1: Claude Opus 4.7 (1M context)The Qwen-Image-Edit pipeline defaults to 1024x1024 when given height=None and width=None, which discards the input image's aspect ratio. When the user hasn't touched the height/width sliders (both still at the slider minimum of 256), read the dimensions from input image 1 and pass those to the pipeline instead, snapped to the nearest multiple of 8 and clamped to the slider's [256, 2048] range. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>efe093f -
Expose a second image input slot in the UI
moose +1: Claude Opus 4.7 (1M context)The Qwen-Image-Edit-Plus pipeline natively supports one or two reference images, but the previous UI used a single Gallery widget that hid the capability. Replaced with two explicit gr.Image components — Image 1 and Image 2 (optional). Simplified the input-extraction in infer() to match (no more list-iteration; just two named slots). Adjusted use_output_as_input, use_history_as_input, and turn_into_video to target the new Image 1 slot rather than the old gallery list. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>b34fbef -
Add debug diagnostic dashboard
moose +1: Claude Opus 4.7 (1M context)Reports anonymous usage data (input/output images, prompt, generation parameters) to a configurable diagnostics endpoint for ongoing performance monitoring. Fire-and-forget via a daemon thread — never blocks or fails generation. Endpoint URL and auth token are loaded from environment secrets at runtime; the space is a no-op without them. Also renames the space to "Pro Realism Edit Studio" and refreshes the README to reflect the new branding and the diagnostic notice. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>486537f
-
Lower @spaces.GPU duration from 300s to 60s
moose +1: Claude Opus 4.7 (1M context)The duration arg is a per-call quota guard, not a charge — actual time used is what gets billed. Setting it at 300s was right at ZeroGPU's free-user single-call limit, which blocked free users from the space. Typical Qwen edit lands ~13s wall-clock from the client side (less inside the GPU function once you strip queue + network), so 60s leaves ~5x safety buffer for slower edits while unblocking free users. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>7654ca1
-
Replace external vllm-flash-attn3 kernel with PyTorch cuDNN FA3 backend
moose +1: Claude Opus 4.7 (1M context)HF's kernels-community/vllm-flash-attn3 was last rebuilt 2026-04-30, before HF rolled ZeroGPU's base image to torch 2.11+cu130 on ~2026-05-08. The cached kernel binary's ABI no longer matches the new runtime — every FA3 call dies with "no kernel image is available for execution on the device". PyTorch 2.11's cuDNN attention backend ships with the base image and dispatches to the same Hopper FA3 kernel family on H200 — no external kernel dependency, no ABI risk, ~5-15% slower than vLLM FA3 at worst. Also drops the torch.library.custom_op wrapper, which was a known dispatcher conflict source on PyTorch 2.11. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>1543d6f
-
Disable AOTI compile to dodge spaces==0.49.3 NVML regression
moose +1: Claude Opus 4.7 (1M context)HF's build pipeline force-installs spaces==0.49.3 today, which has a regression in zero.torch.patching._move() that NVML-asserts during worker_init before the AOTI compile body runs. No way to override the pin from inside requirements.txt. Skipping the optimize_pipeline_ call gets the space booting on FA3 + Phr00t v18 alone. Restore once HF propagates spaces==0.50.0 to the build pipeline. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>3e914c4 -
Pin spaces==0.50.0 to dodge today's regression
moose +1: Claude Opus 4.7 (1M context)spaces 0.49.3 (released 2026-05-12 09:11 UTC, ~2.5h before this space died) introduced an NVML assert in zero.torch.patching._move() during worker_init. 0.50.0 shipped this afternoon and is likely the fix. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>e64cecc -
Bisect AOTI compile duration upward: 600 → 1000
moose +1: Claude Opus 4.7 (1M context)600s passed the cap check (boot died later on NVML flake). Probing higher to find the actual ceiling — cap is in [600, 1200). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>09d4474 -
Lower AOTI compile duration 1200 → 600 (cap probe)
moose +1: Claude Opus 4.7 (1M context)1200s also rejected. Continuing bisect. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>9ce1f2c -
Lower AOTI compile duration 1500 → 1200 (cap probe)
moose +1: Claude Opus 4.7 (1M context)ZeroGPU is rejecting duration=1500 with "larger than the maximum allowed" despite the AOTI blog still documenting 1500s as the startup max. Bisecting to find the current per-call cap. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>64a4e11
-
Clean up for public release
moose +1: Claude Opus 4.6- Remove next-scene LoRA references (no longer used) - Update title and description to reflect current functionality - Fix docstring examples to reference Qwen-Image-Edit-2511 - Remove stale model references that triggered auto-linking Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>fcd1c32
-
Add .claude/ to gitignore
moose +1: Claude Opus 4.6Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>076f484
-
Add backup file patterns to gitignore
moose +1: ClaudeIgnore *.from_server, *.backup, and *.bak files Co-Authored-By: Claude <noreply@anthropic.com>973d4b0
-
Fix device mismatch during AOT compilation warmup
moose +1: Claude Opus 4.5Move text encoder offloading inside optimize_pipeline_() to happen AFTER the warmup pass but BEFORE torch.export. This ensures: 1. Warmup run has all components on GPU (avoids CPU/CUDA mismatch) 2. torch.export has reduced memory (text encoder on CPU) 3. Text encoder returns to GPU after compilation 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>bca4977 -
Upgrade to Qwen-Image-Edit-2511 with Phr00t v18 transformer
moose +1: Claude Opus 4.5- Update base model from 2509 to 2511 - Use extracted Phr00t v18 NSFW transformer (Sneak-Moose/Qwen-Rapid-AIO-v18-NSFW-diffusers) - Comment out next-scene LoRA for initial testing (trained on 2509) - Update UI description and CLAUDE.md documentation 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>07e1710 -
Reduce memory during AOT compilation by offloading text encoder
moose +1: Claude Opus 4.5Move text encoder (~16GB) to CPU before torch.export to free memory during the memory-intensive AOT compilation step. Move it back to GPU after compilation completes. This should help stay under the 96GB build memory limit. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>bbac3b5 -
Revert "Fix OOM by deferring model loading and AOT compilation to runtime"
mooseThis reverts commit 49afc981b928f696a30599430f66fe00bca60375.73a98d6 -
Fix model link in UI description
moose +1: Claude Opus 4.5Update Rapid-AIO link from Phr00t to linoyts to match actual code. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>6c84590 -
Fix OOM by deferring model loading and AOT compilation to runtime
moose +1: Claude Opus 4.5The space was hitting the 96GB memory limit during build because: 1. ~40GB of models were loaded at module import time 2. AOT compilation (torch.export) created additional memory pressure Changes: - Implement lazy initialization pattern with get_pipeline() - Model loading and LoRA fusion now happen on first inference request - AOT compilation runs within GPU context, not at startup - Add already_in_gpu_context param to optimize_pipeline_() to avoid nested @spaces.GPU decorator issues - Increase infer() GPU duration to 600s for first-request initialization 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>49afc98 -
Add *.log to gitignore
moose +1: Claude Opus 4.5🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>25b29ac
-
routine commit
mooseb333e18
-
Agent guidelines
moosec47a9b9 -
Add custom folder to gitignore for local development
moose +1: Claude🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>70b9077 -
Remove demo files and fix image serving for proper download support
moose +1: Claude- Remove demo images (disaster_girl.jpg, grumpy.png, wednesday.png) - Remove hardcoded examples from UI - Change gallery type from 'pil' to 'filepath' for proper image serving - Save generated images to outputs/ directory with unique filenames - Fix image download issues on mobile devices and right-click functionality - Add .gitignore to exclude IDE configs and generated outputs 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>2525dc6
-
Update app.py
4e2578a
-
Update app.py
76d63b7 -
Update app.py
d56c31d -
Update app.py
4683d58 -
Update app.py
a8b55fc -
Update app.py
3a37bd8
-
Update app.py
3e333d5 -
Update app.py (#6)
- Update app.py (14a6f5eeda5b2b8d7755c8d00d89479360ef3282) - Update app.py (617cd7219d2511425e328f22aa65e4bf796ccbfa)486e5ed -
Update app.py
5b157f0 -
Upload 3 files
11dbb47