AI setup

Lumibelle works without any AI provider. Connect one when you want writing assistance, reference images or video generation. Lumibelle never downloads models, installs nodes, or switches models or providers on your behalf.

AI settings

Open AI settings from Script or Assets. Connections, Text models, Image models and Video models list providers, workflows or sections in a sidebar, each with its own settings. The model tabs end with a Defaults entry for settings that apply to all of them, such as the global text model, the default image workflow, request timeouts and local media tools. Text models contains a searchable catalog with provider and starred filters; LoRAs holds the LoRA library. Refresh a provider explicitly to discover models. Opening settings never generates anything.

Star models to build a shortlist shared by every project. ComfyUI models must pass a test against the current server before starring; OpenRouter requires a valid key and a successful catalog check. Stars, global-default changes and successful test results save immediately. Each setup form has its own Save and Cancel.

Project text model

All text assistance uses one model default per project, set in Project settings → Text assistance or through Set as project default in any model menu. New projects start with Use global default. The model chip inside assistance opens configured choices and reasoning options. A temporary override applies to one successfully queued request, and retries keep their captured inputs. Missing or incompatible models never trigger substitution.

Text-assistance composers share a compact dialog that becomes a full-width sheet on phones. Closing keeps draft instructions and model overrides.

ComfyUI

ComfyUI provides local text generation, reference images and image editing.

1. Start an up-to-date ComfyUI with the built-in CLIPLoader, TextGenerate and PreviewAny nodes.
2. Install a text-generation encoder supported by those nodes. The initial model is gemma4_e4b_it_fp8_scaled.safetensors when installed. Lumibelle lists every exact filename that CLIPLoader advertises.
3. For reference images, install a Krea 2 Turbo diffusion model, a Qwen3-VL 4B text encoder and qwen_image_vae.safetensors. The defaults prefer krea2_turbo_int8_convrot.safetensors, qwen3vl_4b_fp8_scaled.safetensors and qwen_image_vae.safetensors when detected.
4. For image editing, install the current comfyui-krea2edit and comfyui-tooling-nodes node packs, restart ComfyUI, and place krea2_identity_edit_v1_2.safetensors in the ComfyUI LoRA directory. Missing edit dependencies disable editing without disabling image creation or manual asset work.
5. Enter the server URL (default http://127.0.0.1:8188) in Connections, check it, and save. In Text models, choose Refresh ComfyUI, then Details & test on a model. Test selected model · 256 tokens runs the normal workflow with a fixed prompt. A successful result saves immediately and enables the star. Refresh and save the image and Krea 2 Edit LoRA files separately in Image models.

The optional Advanced model test accepts a custom message and output limit so you can inspect model behavior; its prompt and response are never saved. Before a model test, Lumibelle checks that ComfyUI has no running or pending work and asks it to free memory; a busy queue blocks the test instead of disturbing another job. Use Test again after replacing a model file without changing its name.

Each successful test records an observed token rate and, when ComfyUI reports it, VRAM use. The latest five observations per model are kept to help compare models on the same machine.

Reference generation follows the official Krea 2 Turbo workflow with the exact prompt you entered. Lumibelle copies each completed image into the project.

Progress and cancellation

ComfyUI queues Lumibelle’s work alongside other jobs. Lumibelle shows the current workflow stage, elapsed time, generated tokens or sampling steps, and an approximate time remaining when ComfyUI supplies measurable progress. The timeout includes time spent queued. Cancel targets only that job; Lumibelle never sends a global interrupt or clears the queue.

FLUX.2 Klein 9B KV

FLUX.2 Klein 9B KV is also available for creation and multi-reference editing. In AI settings → Image models, select FLUX.2 Klein 9B KV, refresh the catalog, and save; choose it as the default workflow under Defaults if you want it for new requests. Each workflow keeps its own file selections, and Assets has an Image workflow selector for the current run.

Install these files in the corresponding ComfyUI model directories (subfolders are supported):

• Diffusion model: flux-2-klein-9b-kv-fp8.safetensors in models/diffusion_models, from Black Forest Labs.
• Text encoder: qwen_3_8b_fp8mixed.safetensors in models/text_encoders, from Comfy-Org’s Klein 9B files.
• VAE: flux2-vae.safetensors in models/vae, from Comfy-Org’s FLUX.2 files.

Use a current ComfyUI with FluxKVCache, Flux2Scheduler, EmptyFlux2LatentImage, ReferenceLatent and the custom sampler nodes. Editing also uses ETN_LoadImageBase64 from the tooling node pack. A missing KV model does not fall back to standard Klein, a base checkpoint, 4B or Krea.

LoRAs

In AI settings → LoRAs, refresh installed files and register an exact file with a friendly name, workflow assignment, default strength and optional trigger text. Registrations belong to their ComfyUI server; they organize installed files without downloading them or testing compatibility. Krea’s required identity-edit LoRA stays separate.

Add comma-separated Tags to LoRA registrations, then search or filter the library by tag. In a project’s Settings tab, LoRA visibility applies to every asset and both Create and Edit:

• Empty filters show all LoRAs.
• Only show these tags matches any listed tag and hides untagged LoRAs.
• Hide these tags always wins, even when an allowed tag also matches. For example, hide nsfw in an SFW project.

An already selected, enabled LoRA that the project now excludes blocks generation until it is disabled, removed or allowed again. Each batch checks the saved project rules when it starts; changes never rewrite earlier images or their recorded LoRAs. To use LoRAs on an asset, see Assets studio.

OpenRouter

1. Enter your own OpenRouter API key in Connections, check it, and choose Save key.
2. In Text models, choose Refresh OpenRouter, search for a model, and star it. Optionally choose Set default.
3. Select the starred model through the model chip inside Assist.

Checking the key and refreshing models generate no text. Requests stream text into the assistant; a model must be explicitly selected, and Lumibelle does not retry paid generation requests automatically.

Default advanced settings are temperature 0.7, maximum output 2,048 tokens and a 300-second timeout. Authentication, credit, rate-limit, context, model, connection and execution errors are shown. Partial text and instructions remain available after a failure or cancellation.

Claude Code

Lumibelle can use an installed Claude Code CLI for text assistance, including optional image inspection. It runs with the account you set the CLI up with, and never signs in, reads credentials or modifies the CLI. Claude Code cannot generate images.

1. Install Claude Code 2.1.259 or later (see Claude Code setup), open a new terminal and run claude once to sign in, as the same operating-system user that runs Lumibelle. Connections → Claude Code → Install the Claude Code CLI shows the command for your system.
2. Open AI settings → Connections → Claude Code, enable the provider, and Check connection. Leave the executable path empty for detection, or select the installed executable.
3. Save the connection. The check shows the provider, account and sign-in method that requests will use; it generates no content.
4. In Text models, star the models you want and optionally set Default Claude Code reasoning effort.

Claude Code does not report remaining usage, so Lumibelle shows no allowance meter. When a request ends because of a usage or rate limit, it needs attention and the Claude Code queue pauses until you choose Resume.

Codex

Lumibelle can use an installed Codex CLI for text assistance and image Create/Edit, with the account the CLI is set up with.

1. Install Codex CLI 0.153.4 or later (see Codex CLI), then run codex login as the same operating-system user that runs Lumibelle. Connections → Codex → Install the Codex CLI shows the command for your system.
2. Open AI settings → Connections → Codex, enable the provider, and Check connection.
3. Save the connection. The check reports the CLI version, account and, for ChatGPT sign-ins, the plan allowance; it generates no content.
4. In Text models, refresh Codex, star the models you want and optionally set Default Codex reasoning effort.
5. For images, select Codex Images in Image models, refresh, choose the Codex agent model and Agent reasoning effort, and save. Choose it under Image models → Defaults to use it for new image requests.

Codex image edits support one base plus up to seven additional references, each with its own crop. For ChatGPT sign-ins, the plan allowance is shown for information only. When Codex reports an exhausted limit, that request needs attention and the Codex queue pauses until you choose Resume.