Cheapsfield v0.1.0 — 100% local, open-weights creative studio

Cheapsfield v0.1.0

Cheapsfield is a 100% local, 100% open-weights creative studio for Windows: Higgsfield-style apps and a node canvas, built on ComfyUI and driven headless by the Comfy SDK.

Everything runs on your own NVIDIA GPU with open-weights models. No cloud, no API keys, no credits, no uploads.

Built for the Comfy Open Source Higgsfield challenge (Category: Open Weights Only · Bonus: VFX Workflow Built with LTX).

---

Highlights

• Spot Edit (new): mark the photo with an Area, a Pin, an Arrow or a Sketch, write what to change at each mark and optionally attach a reference image (REF1/REF2) to a mark. Only the marked area is edited and pasted back with a soft mask, so the rest of the photo stays pixel-identical. Works with Qwen Image Edit 2511, Qwen Image 2.1 and FLUX 3 Klein (core nodes only: ImageCrop, ImageScale, ImageCompositeMasked).
• Character Swap: replace the person (or the scene) in a video with a reference photo, keeping the motion, timing and camera (MiniMax H3 ref2va, video to video).
• Create Influencer: build a fictional digital influencer from a short form (look, age, hair, outfit), ready for Character Swap.
• Reveal Video (VFX): before → after video with native sound (sketch → product shot, empty → furnished, day → night), up to 1080p with LTX 2.5 HD.
• True Line: sketch, 3D viewport or clay model → photorealistic image with the exact same framing.
• 16 one-screen apps that write the prompt for you, with a Subject: Architecture / Product switch in 11 of them. Also: Styled Plan, Light & Hour, Reference Mood, Orbit (360° rotation, tilt and zoom), Furnish from Moodboard, Object in Scene, Text Retouch, Render Finisher, Map to Drone, Ink Sketch, Beyond the Frame and Upscale.
• Flow: a node canvas with Generate image, Edit image, Generate video, Spot Edit and every app as a node.
• Embedded recipes: every generated PNG stores its recipe, so you can rebuild the node that made it in one click.
• First-run setup: finds your ComfyUI and models, prepares the Comfy SDK bridge, picks the GPU and downloads missing models (with resume).
• Core ComfyUI nodes only: no custom nodes and no partner API nodes.

---

System Requirements

• OS: Windows 10 / 11 (64-bit).
• GPU: NVIDIA with at least 12 GB VRAM (16 GB+ recommended for video).
• RAM: 32 GB recommended.
• ComfyUI: version 0.41.0 or newer (Portable, ComfyUI-Easy-Install or a Git clone with its Python environment).
• Storage: from ~17 GB (one model) up to ~205 GB (every model).

---

Install

1. Download and extract Cheapsfield-0.1.0-windows-x64.zip (for example to C:\Apps\Cheapsfield).
2. Run Cheapsfield.exe.
3. The setup wizard guides you through:
   • choosing your library folder (SQLite, never in AppData);
   • pointing to your ComfyUI folder and clicking Prepare the Comfy SDK (~20 s);
   • choosing the GPU;
   • checking your models or downloading the missing ones from Hugging Face.
4. Open an app or a Flow, drop an image and click Generate.

SHA-256 of Cheapsfield-0.1.0-windows-x64.zip:
876cc6789570948d68a2269e15dd050a966c014c16625b44562efaf9402f4ce2

---

Open-Weights Models & Licensing

• Images and editing: Qwen Image 2.1, Qwen Image Edit 2511, FLUX 3 Klein (base and distilled, with the prompt upsampler), Krea 2, FLUX.1 Fill.
• Video and VFX: LTX 2.5 (2-stage 1080p HD) and MiniMax H3 (native audio; ref2va + turbo LoRA for Character Swap).
• Upscaling: 4x-UltraSharp.

The Cheapsfield source code is under the Apache-2.0 license. Each model keeps its own license; read docs/MODELOS.md before commercial use.

---

Known Limitations

• The first generation after start takes longer while the weights load into VRAM.
• 1080p video with LTX 2.5 needs ~16 GB VRAM; with 12 GB it offloads to system RAM and renders slower.
• Spot Edit with only a Pin or an Arrow (no Area) lets the model redraw the whole photo; add an Area to lock the rest.
• Windows x64 with an NVIDIA GPU only, for now.