FrameLoom brings LLM chat, reference-based editing, and Qwen-Image-2.1 generation together in one workspace. Discuss an idea, assign reference images, refine the composition, and prepare your prompt step by step - then generate when you’re ready.
Four integrated panels - Models & LoRA, LLM / MANUAL, Finishing, and Output - keep the workflow compact, with a large image preview beside your controls.
⚠️ Content Disclaimer
FrameLoom is a creative tool. You are responsible for your prompts, reference materials, generated content, and how that content is used, shared, or published. The author does not review or endorse user-generated content.
Please respect third-party rights and follow the licenses and terms of the models, dependencies, and API providers you use.
💡 My Recommendation
I recommend using larger models through OpenRouter, especially for complex scenes, iterative editing, and the character builder. These modes benefit from a model that can follow detailed instructions and keep track of many requirements.
You can also connect a local LLM. The experience will depend on the model you choose and the memory available on your system.
I recommend using DeepSeek V4.1 Flash 1017. Based on my statistics (1,000 requests in a mixed mode—text-only and text-with-images), each request costs $0.001542857
You can use OpenRouter for free by creating an account and generating an API key. Then, either select a model from the list of free options or enter 'openrouter/free' into the 'model' field.
✨ Key Features
Chat inside your workflow. Discuss ideas and prepare prompts independently of image generation.
Cloud and local assistants. Connect to OpenRouter, or use LM Studio and Ollama running on the same computer as ComfyUI.
LLM or MANUAL control. Work with an assistant or write your prompts directly. The positive prompt remains editable.
Instructions for the selected task. The built-in system combines shared rules with instructions relevant to your chosen task and inputs.
Review before generating. Inspect the prompt and proposed settings. Automatic reply application can be disabled.
Custom system prompts. Edit the built-in instructions, add a supplement, replace them entirely, and save named presets.
Separate chat and generation. Generating another image reuses the applied prompt without making another chat request.
🧠 14 Image Tasks
Choose the task in Parameters before sending your request:
Generate a new image or compose a scene from references.
Edit an existing image.
Create typography and layouts with requested text.
Generate transparent images, remove backgrounds, or edit RGBA content.
Inpaint selected regions or extend the canvas.
Work with identity, objects, clothing, pose, and style references.
The selected task stays under your control. If the conversation requires another task, change it in the interface before continuing.
🖼 Images & References
Upload images, drag them into the reference area, paste them into the chat composer with Ctrl+V, or reuse a generated result.
Assign each image its contribution: identity, face, hair, body build, clothing, pose, style, composition, background, or source canvas. Add individual notes, reorder references, and choose which images are used by Qwen.
The workflow supports up to 16 encoder image slots, including a separately routed reference mask.
Use the crop editor to isolate a useful detail while retaining the original for restoration. References, roles, notes, and chat history are saved with the workflow. New chat clears the working context, with an in-session undo available.
Image analysis requires a vision-capable LLM and Send images enabled.
🎭 Masks, Transparency & Structural Control
Edit selected regions using reference masks, latent noise masks, or optional Union inpainting.
Use white areas for editing and black areas for protection, or select transparent areas through the mask’s alpha channel.
Paint annotations and expand the canvas in the image editor.
Work with PNG alpha for transparent generation, background extraction, and transparent edits.
Use optional Union ControlNet with Canny, depth, grayscale, HED, line art, MLSD, pose, and scribble maps.
Prepared maps and photograph-based preprocessing are separate options. Compatible preprocessors and Union weights must be installed for the branches you use.
➕ Positive & Negative Prompts
The Negative OFF / ON switch works alongside LLM and MANUAL modes.
In LLM mode, the assistant prepares a coordinated positive/negative pair. In MANUAL mode, you can edit both fields yourself. Turning Negative OFF preserves its draft and restores the previous CFG and step settings.
At CFG 1, standard sampling ignores negative guidance. The Turbo LoRA negative-mode setting at CFG 1.1 is experimental.
📚 LoRA Library
Browse compatible local LoRAs with search and filtering.
Save descriptions and exact trigger words for reuse.
Pass information about selected LoRAs to the assistant.
Adjust strengths and ordering individually.
Refresh the catalogue after adding files.
📐 Canvas & Generation Controls
Choose an aspect-ratio preset, enter custom dimensions, or swap width and height. Lock the canvas when the assistant should adapt the composition to your selected size.
A separate permission controls whether the LLM may change ControlNet settings.
Select Quality, Turbo LoRA, full Turbo checkpoint, or Custom settings profiles. Sampling controls include steps, CFG, sampler, scheduler, denoise, batch size, and seed.
⚙️ Optional Processing
A second sampling pass.
Face refinement with separate prompt and sampling controls.
Model-based neural upscaling.
Tiled VAE decoding and prefix KV cache options.
Compatible GGUF loading through ComfyUI-GGUF.
Separate local prompt enhancement with Qwen PE-T2I / PE-I2I models.
Optional branches require their corresponding weights and dependencies. The public example starts with optional processing disabled.
💾 Preview & Saving
Preview results in the Output panel and reuse them as references.
Save PNG files to ComfyUI’s output directory or another existing folder on the backend computer. Choose a filename/subfolder prefix and whether to embed prompt and workflow metadata.
The native folder picker currently supports Linux with Zenity. Other setups can use the destination path field.
API credentials are stored separately in browser storage and are excluded from workflow JSON.
🌐 Interface & Prompt Language
The workflow uses an English interface.
Conversation language, generation-prompt language, and text rendered inside an image are handled separately. Prompts default to English; requested lettering can retain its original language and script.
💡 Before You Generate
Review complex reference transfers, pose instructions, masks, and proposed settings before generating. The assistant can still misread an image or introduce unwanted details. Prompt instructions cannot guarantee exact identity, anatomy, geometry, or preservation.
📦 Installation
The ZIP includes the four-node workflow JSON, custom node pack, README, and separate Windows and Linux installation guides. Model weights are not included.
Release: 4.1.0rc1.
The custom node code is released under the MIT license. Model weights and third-party components retain their own licenses.
The workflow has been developed and tested on Linux. Windows installation instructions are included; Windows runtime compatibility has not yet been independently verified.