Spaces:
Paused
Paused
Commit History
-
Fix: save Kling Motion video to settings.paths.output_dir (matches serve endpoint)
5a8b9c0 -
Fix: reuse provider http client for Kling Motion, fix indentation
27e64d0 -
Fix: auto-resize character image before upload for Kling Motion
a97bced -
Fix: move character_orientation into function signature (was in body)
19c97d8 -
Fix: add required character_orientation param to Kling Motion API call
36371b1 -
Fix: Kling Motion set status to 'completed' to match frontend poll check
49af4b6 -
Kling Motion: extend poll timeout to 25 min (300x5s)
455d927 -
Fix: correct kling-motion route path (remove double prefix)
1970317 -
Add Kling Motion Control: character image + driving video via WaveSpeed
b02f80a
-
Fix: add wavespeed to requirements, make SDK import optional
4edd939 -
Sync all local changes: video routes, pod management, wavespeed, UI updates
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>e808ae1
-
Add LoRA strength slider, auto-set FLUX.2 defaults (CFG 2), show advanced settings for pod
4340a68 -
Fix: request.prompt not request.positive_prompt
38899e5 -
Fix catalog: use insert_image with correct params, store prompt info in job
9ff31f1 -
Add /pod/jobs/{id}/image endpoint, show generated image directly in UI
302a193 -
Add pod generation progress polling with live status in UI
79598c0 -
Increase gen timeout to 900s, add progress logging
d812254 -
Fix detached training launch: use setsid, close FDs, add debug logging
f072956 -
Add bulk caption paste feature to training UI
1dc4764 -
Fix nohup quoting: write train command to script file first
The train_cmd contains single quotes (optimizer args) which broke nohup bash -c '...'. Now writes to /tmp/train.sh via heredoc then runs nohup on the script file. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>85988f2 -
Fix coroutine .strip() bug — wrap await in parentheses
await self._ssh_exec(...).strip() calls .strip() on the coroutine, not the result. Fixed all 11 occurrences to (await ...).strip(). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>ce5bf6b -
Auto-reconnect to running training pods after server restart
On startup, checks DB for in-progress training jobs, verifies their RunPod pods are still running, and reconnects SSH to resume log monitoring. Handles both still-running and already-completed cases. Copies LoRA to volume and terminates pod on completion. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>1c46fa4 -
Run training in detached nohup process (survives SSH/server disconnect)
Training now runs via nohup with output to /tmp/training.log. Monitor reads log file instead of SSH channel stdout. Safe against HF Space rebuilds, SSH drops, and browser closes. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>b4ceba9 -
Fix event loop blocking: make all SSH/SFTP calls async
All _ssh_exec calls now use asyncio.to_thread to avoid blocking the web server during training setup. SFTP put/get also wrapped. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>3aa914c -
Add debug logging for DB path on startup
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>b341f22 -
Skip local LoRA download on HF Spaces, clean up training uploads
LoRA is saved to RunPod network volume (ready for generation). Local download skipped on HF to avoid filling 1GB storage limit. Training upload images cleaned up after upload to RunPod. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>2d79152 -
Fix HF Spaces startup: ensure DB directory exists before engine init
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>c231a93 -
Switch training UI from epochs to max steps (default 1500)
- Replace Epochs field with Max Steps (1500 default, 1500-2000 recommended) - Replace Save Every N Epochs with Save Every N Steps (500 default) - Persist pod state to disk so it survives server restarts - Increase generation polling timeout to 600s Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>01a9c08 -
Add pod generation with FLUX.2, persistent state, training improvements
- Pod management: network volume support, HTTPS proxy URLs, fire-and-forget SSH - FLUX.2 workflow: separate UNETLoader + CLIPLoader + VAELoader with Comfy-Org text encoder - Persist pod state to disk (survives server restarts) - Training: persistent job tracking in DB, live log streaming - Remove tracked __pycache__ files, add .gitignore Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>27fea48
-
Clean up UI: smaller controls, better spacing
- Narrower controls panel (340px) - Smaller form elements (6px padding, 12px font) - Compact chips (4px padding, square corners) - Smaller drop zones (16px padding, 100px min-height) - Section dividers with border-top - Inline character/template selectors in 2-column grid - Shorter labels and placeholder text - Compact reference image previews (70px max) - Better visual hierarchy with cleaner spacing Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>b051fff -
Make wavespeed SDK optional, use httpx fallback
- Import wavespeed SDK conditionally (try/except) - Fall back to direct HTTP API via httpx when SDK unavailable - Remove wavespeed from requirements (may cause install hangs) - Update is_available() and submit_generation() to handle both cases Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>88af028 -
Fix startup: add runpod to requirements, make import optional
- Add runpod>=1.6.0 to requirements.txt - Remove higgsfield-client (not needed, SDK doesn't exist) - Make runpod import optional in routes_video.py Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>5402b16 -
Add multi-reference image support + fix missing python-dotenv
- Add python-dotenv to requirements.txt (was missing, causing import error) - Add multi-reference image support for img2img (character + pose reference) - Add pose/style drop zone to UI with side-by-side layout - Add multi-ref models: SeeDream Sequential, Kling O1, Qwen Multi-Angle - Update wavespeed_provider to handle multiple image uploads Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>f723987 -
Major UI cleanup and improvements
- Fixed duplicate model selector issue in img2video mode - Replaced width/height with simple Aspect Ratio dropdown - Collapsed advanced settings (steps, cfg) by default - Reorganized model dropdowns with clearer categories: - Recommended, NSFW Friendly, Fast, Other - Simplified video model selector (removed redundant backend chips) - Added pricing hints to video models - Cleaner labels and descriptions throughout - Added updateDimensions() function for aspect ratio Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>21323c4 -
Add detailed logging for video generation debugging
- Log video generation request details - Log image upload status - Log WaveSpeed API responses - Update job status messages for better UI feedback Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>1ebf6af -
Add Higgsfield provider for Kling 3.0, Sora 2, Veo 3.1
- Created HiggsFieldProvider using higgsfield-client SDK - Added Kling 3.0 Pro and standard models for video generation - Added Sora 2 and Veo 3.1 via Higgsfield platform - Updated UI with premium model options - Added higgsfield-client to requirements Requires HIGGSFIELD_API_KEY or HF_API_KEY env var to use. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>a4583f2 -
Add video async polling + Higgsfield models via WaveSpeed
- Fixed video generation async polling (same issue as img2img) - Added Higgsfield Soul (image edit) and DoP (video) via WaveSpeed API - Updated UI dropdowns with Higgsfield options Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>fc4811e -
Fix async polling - check urls.get BEFORE outputs check
Reordered conditions to check for async response first. Added debug logging for response data keys. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>f1f21b9 -
Fix WaveSpeed async job handling - add polling for outputs
WaveSpeed returns async responses with urls.get for some models. Added _poll_for_result method to poll until outputs are ready. Also added _resolve_video_model helper. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>508b150 -
Add WAN 2.2 and Qwen edit models to image-to-image
- Added WAN 2.2 Edit to EDIT_MODEL_MAP - Added Qwen Edit models (edit-plus-lora, edit-multiple-angles, layered) - Added WAN 2.6 I2V Pro and Flash variants - Added WAN 2.2 I2V 1080p and 720p variants - Updated UI dropdowns for both edit and video modes Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>7c6b44e -
Fix WaveSpeed model IDs - use real API model paths
- Updated MODEL_MAP with real WaveSpeed text-to-image model IDs - Updated VIDEO_MODEL_MAP with real image-to-video model IDs - Updated EDIT_MODEL_MAP with real edit model IDs - Fixed UI dropdowns to match real available models - Removed fake NSFW models that don't exist on WaveSpeed - Added real models: GPT Image, WAN, Kling, Dreamina, Vidu Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>68e0e77 -
Fix: Add job tracking to img2img cloud generation
136976f -
Add job tracking, status display, and cancel button
Features: - Job status tracking for cloud generations (txt2img, img2img) - Real-time status messages during generation - Cancel button to stop running jobs - Job status API endpoints (GET /generate/jobs, GET /generate/jobs/{id}, POST /generate/jobs/{id}/cancel) - Shows detailed error messages when generation fails Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>08543e5 -
Add NSFW-friendly models to Image-to-Image edit
Edit models added: - Qwen 2.5 VL Edit (less restrictive) - Pony v7 Edit (anime/NSFW) - Illustrious XL Edit (anime) - FLUX Dev Uncensored Edit Organized both txt2img and img2img dropdowns with NSFW Friendly sections Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>7b383a1 -
Add Qwen and NSFW-friendly models
Text-to-Image: - Qwen 2.5 VL 72B, Qwen 2 VL 72B (less restrictive) - Pony Diffusion v6/v7 (anime/NSFW friendly) - Illustrious XL (anime) - FLUX Dev Uncensored - SDXL Lightning, SDXL Turbo Organized dropdown with "NSFW Friendly" section for easy access Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>4b9b1b8 -
Add API log panel for debugging generation requests
- Added collapsible API log panel below preview - Logs all API calls with method, URL, status code - Shows error details when requests fail - Wraps fetch to automatically capture all API activity - Helps debug cancelled/failed generations Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>4deeec0 -
Add more edit models and fix img2video mode display bug
- Added FLUX Pro/Dev Edit, Dreamina Edit, Qwen Edit to img2img models - Fixed bug where img2video section didn't show (backend-section wrapper) - Organized edit model dropdown with optgroups Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>7634d60 -
Add video model selector with cloud/pod backend support
- Added video model dropdown for img2video mode (WAN I2V, Kling, Veo, Seedance, Sora) - Added backend selector for video generation (Cloud API vs RunPod) - Added cloud video generation endpoint using WaveSpeed API - Added more text-to-image models (FLUX Pro, WAN 2.6, Dreamina, Qwen) - Added more video models (Kling O1/O3/Pro, Veo 3/3.1, Sora 2, Seedance) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>7b5f7c4 -
Add password protection
4b21f3a -
Fix database directory creation at startup
739bb39