ComfyUI Node
Atlas Assess Image 🧭
A ComfyUI node in Atlas/09 · QA & Gates with 11 inputs and 16 outputs.
Atlas Assess Image 🧭
- image
- image
- report
- settings_json
- sam_prompt_sky
- sam_prompt_far
- sam_prompt_bg
- sam_prompt_mid
- sam_prompt_fg
- geom_far
- geom_bg
- geom_mid
- geom_fg
- band_far
- band_bg
- band_mid
- band_fg
◄providerollama►
◄model►
◄base_url►
◄extra_instructions►
◄proceedfalse►
◄approved_for►
◄api_key►
◄offload_modelfalse►
◄auto_continuetrue►
◄offload_ttl_s2►
CategoryAtlas/09 · QA & Gates
Inputs (11)
| Name | Type | Default | Description |
|---|---|---|---|
| image | IMAGE | — | |
| provideropt | COMBO | ollama | VLM backend. ollama/lmstudio/llamacpp are local servers; 'openai' is any OpenAI-compatible CLOUD endpoint (api.openai.com by default, OpenRouter etc. via base_url) for users without local models — needs api_key. Blank model/base_url use each provider's own defaults (same conventions as AtlasVLMScaleCues). |
| modelopt | STRING | Vision model id; blank = provider default (ollama: gemma3:4b). | |
| base_urlopt | STRING | — | |
| extra_instructionsopt | STRING | Optional artist notes appended to the assessment request — e.g. 'the camera move is a slow dolly-in on the tower'. The VLM tailors band/patch advice to the intended move. | |
| proceedopt | BOOLEAN | false | OFF = the image output is paused (downstream skipped) so you can read the report and apply settings first. Turn ON (or click ▶ Continue Workflow) and re-queue to run the full pipeline. A ▶ Continue click approves THIS image only (see approved_for); a manual toggle here is an unconditional override. |
| approved_foropt | STRING | Managed by ▶ Continue Workflow: the fingerprint of the image the current proceed=True was approved for. When the input image changes, the gate re-arms automatically instead of running a stale approval. Leave empty when toggling proceed by hand (empty = unconditional). | |
| api_keyopt | STRING | API key for the 'openai' cloud provider (ignored by local providers). SAVED INTO THE WORKFLOW FILE — prefer leaving this blank and setting the OPENAI_API_KEY environment variable so shared workflows never carry your key. | |
| offload_modelopt | BOOLEAN | false | Free the VLM's VRAM after a SUCCESSFUL assessment so the heavy pipeline (depth/SAM/LaMa) doesn't fight it for memory — the assessment is cached per image, so ▶ Continue never reloads the model. Per provider: ollama = keep_alive:0 (clean unload); lmstudio = request ttl for JIT loads + the 'lms' CLI when on PATH for GUI-loaded models; llamacpp = NOT possible (the server owns its model — restart it to free VRAM); openai = nothing local. A failed assessment keeps the model warm for the retry. The report shows what actually happened. |
| auto_continueopt | BOOLEAN | true | ON (default): advisory mode — the assessment runs, its SAM prompts/geometry flow downstream, and the SAME queue continues without a ▶ Continue click; the ✅ solve gate (and the later 📐 patch gate) become the workflow's checkpoints. Turn OFF to restore the hard per-image gate: the image output blocks until ▶ Continue Workflow approves THIS image. |
| offload_ttl_sopt | INT | 21–3600 | LM Studio only, with offload_model ON: idle seconds before the VLM auto-evicts. 2 = free RAM immediately after each assessment. 120-300 = linger so a BATCH of plates skips the ~10-20s reload per image, then free. Reload is automatic (JIT) either way — the next assessment brings it back. |
Outputs (16)
| Name | Type | Description |
|---|---|---|
| image | IMAGE | — |
| report | STRING | — |
| settings_json | STRING | — |
| sam_prompt_sky | STRING | — |
| sam_prompt_far | STRING | — |
| sam_prompt_bg | STRING | — |
| sam_prompt_mid | STRING | — |
| sam_prompt_fg | STRING | — |
| geom_far | STRING | — |
| geom_bg | STRING | — |
| geom_mid | STRING | — |
| geom_fg | STRING | — |
| band_far | STRING | — |
| band_bg | STRING | — |
| band_mid | STRING | — |
| band_fg | STRING | — |