v0.3.0 — FLUX.2 Klein, Wan 2.2 / LTX-Video & Multi-Profile Agent Tooling

🚀 Highlights in v0.3.0

🎬 Open Video Generation Profiles (Wan 2.2 5B TI2V & Lightricks LTX-Video 2B)

• wan-ltx-video Profile: Full support for open-source video generation models on Google Colab L4 GPU:

• Wan 2.2 5B TI2V (official_video_wan2_2_5B_ti2v.json): Text/Image-to-Video generation using Wan22ImageToVideoLatent & WanImageToVideo. Generates smooth 24 fps HD video in ~65–87 seconds.
• Wan 2.1 1.3B T2V (official_text_to_video_wan.json): Lightweight fast text-to-video baseline (prefetch: true).
• Lightricks LTX-Video 2B (official_ltxv_text_to_video.json & official_ltxv_image_to_video.json): High-efficiency diffusion video synthesis.
  • High-Speed Weight Streaming: 19.8 GB core weights stream into memory disk in ~20.6 seconds (peaking at ~518 MB/s) with complete Colab L4 Ada Lovelace optimizations.

🎨 FLUX.2 Klein 4B & Instruction Image Editing (ReferenceLatent)

• flux-klein Profile: Ultra-fast next-generation flow-matching diffusion on L4:

• FLUX.2 Klein 4B Distilled (Apache-2.0): 4-step turbo generation (official_image_flux2_klein_text_to_image.json).
• ReferenceLatent Instruction Editing: Edit images using natural language while locking face identity and composition in ~8 seconds (official_image_flux2_klein_image_edit_4b_distilled.json & _base.json).
• FLUX.1 Schnell fp8: High-speed fallback workflow.
  • Streams 15.4 GB core weights in ~11.5 seconds (at ~616 MB/s).

🤖 Multi-Profile Agent Tooling & MCP Server Upgrades

• Video & Edit MCP Schemas: gate_wake, gate_txt2img, gate_img2img MCP tools and ckit generate now support all 5 built-in profiles (moss-image, flux-klein, wan-ltx-video, sd-latest, qwen-image).
• Aspect Ratio & Multiples Safeguards: Built-in 32-multiple rule enforcement for Wan 2.2 / LTX latent video encoding (e.g. 832×480, 832×544).
• Glama MCP Server Compliance: Full handling of notifications, prompts, resources, and stdio execution for Glama directory indexing.

🌐 Showcase Site & Documentation Overhaul

• Live Showcase: Updated apps/site/ (deployed to Cloudflare Pages: comfy-colab-kit.firetable.tech) with 6-card profile grid and HTML5 live L4 video showcase.
• 2+1 Gallery Layout: README now features an aligned 2-column image session showcase + full-width widescreen cinema video banner in 720p HD.
• Container Image: Published to GitHub Container Registry (ghcr.io/firetable/comfy-colab-kit:v0.3.0 and :latest).

v0.2.0 — AI Agent Scaffolding, Native MCP Img2Img & Colab L4 Acceleration

🚀 Highlights in v0.2.0

🤖 AI Agent Scaffolding & Fork Adaptation (docs/AGENT_SCAFFOLDING.md)

• Build Any ComfyUI Profile with AI: Added a dedicated agent runbook and copy-paste prompt in docs/AGENT_SCAFFOLDING.md so Claude Code, Cursor, Codex, or OpenClaw can automatically inspect Hugging Face models/workflows, configure model_aliases, wire ComfyUI App Mode (app.inputs / app.outputs), and validate profile.yaml via scripts/validate-profile.py.
• Fork for Any Google Colab GPU Workload: Documented the 4 architectural seams (Profile/Plan, packages/colab/setup.py KIT_STAGE stdout protocol, Reverse Proxy probes, and MCP/CLI tools) to let coding agents adapt a fork of this zero-idle gateway for vLLM, Ollama, SD-WebUI Forge, Fooocus, Gradio, Whisper, or custom FastAPI GPU servers.

🎨 Native MCP gate_img2img & Multi-Profile Generation

• New gate_img2img MCP Tool: Available in both local stdio (ckit mcp) and remote HTTP (POST /mcp). Automatically uploads local reference images or HTTP/HTTPS image URLs to ComfyUI (/upload/image) and runs image-to-image generation with configurable denoise (0.05–1.0).
• Profile-Aware Generation: gate_txt2img, gate_img2img, and ckit generate now automatically detect the active profile (moss-image DeBand pipeline or sd-latest Official SD 3.5 Large fp8 pipeline).
• CLI Upgrades (ckit):
  • ckit generate --prompt "..." --image <path_or_url> --denoise 0.75
  • ckit doctor [URL]: One-command diagnostics for local credentials, /_gate/healthz reachability, Cloudflare Access /mcp bypass rules, and Bearer token authentication.
  • ckit validate [path]: Validate custom or community profile directories directly from the CLI.

⚡ Colab L4 Hardware Acceleration Across All Built-in Profiles

• Synced --highvram --reserve-vram 0.2 and Ada Lovelace TensorCore environment optimizations (PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True, CUDA_MODULE_LOADING=LAZY, TORCH_CUDNN_V8_API=1, TORCH_ALLOW_TF32_CUBLAS_OVERRIDE=1) across moss-image, sd-latest, and qwen-image.
• Calibrated cold-start ETA and documentation benchmarks to the measured ~90 s (~1.5 min) end-to-end baseline.

🛡️ Security & Community

• Added .dockerignore to strictly exclude local state, secrets, and custom profiles from Docker build contexts.
• Added CONTRIBUTING.md and GitHub Issue templates for requesting or sharing community profiles.