Server defines 5 tools with explicit schemas and descriptions. Tool naming follows verb_noun conventions (generate_image, get_sd_models, set_sd_model, get_sd_upscalers, upscale_images). All tools have descriptions between 34-81 characters. Input schemas are present and properly structured with JSON Schema types. However, several critical gaps limit production readiness: (1) output schemas are completely undocumented, callers don't know what fields to expect; (2) parameter descriptions lack detail about valid ranges, formats, and error cases; (3) error handling provides no recovery guidance; (4) no tool annotations (destructiveHint, readOnlyHint, idempotentHint) despite clear safety implications; (5) some parameters lack bounds or enum constraints despite accepting free-form strings. The upscale_images tool has a particularly problematic enum definition: resize_mode is declared as string enum ['0', '1'] but parameter descriptions treat it as a number, this type mismatch will confuse LLMs. Overall, the server meets minimum definition requirements but falls short of production-grade clarity and safety patterns.
Generate an image using Stable Diffusion
Get list of available Stable Diffusion models
Get list of available upscaler models
Set the active Stable Diffusion model
Upscale one or more images using Stable Diffusion
Output schemas completely missing for all 5 tools. LLMs have no visibility into what fields to expect in responses. This forces inference and risks downstream tool failures if return structure doesn't match expectations.
Type mismatch in upscale_images: resize_mode schema declares string enum ['0', '1'] but descriptions and environment defaults (SD_RESIZE_MODE parsed as int) treat it as a number. LLM receives contradictory signals.
Sampler and scheduler parameters lack enum constraints. generate_image accepts sampler_name and scheduler_name as free-form strings, inviting LLMs to hallucinate invalid sampler names like 'Quantum Diffusion' or 'Super Euler'.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 60 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Set_sd_model is clearly write/stateful; generate_image and upscale_images are destructive (create files). LLM has no annotation to guide idempotency reasoning or retry safety.
No error recovery guidance. Tool descriptions do not explain what to do if generation fails, timeouts, models are unavailable, or paths are invalid. LLM must guess next steps.
Parameter dependencies undocumented. set_sd_model expects model_name from get_sd_models, and upscale_images expects upscaler_1 from get_sd_upscalers, but descriptions don't state this. LLM may pass arbitrary strings without discovery.
Missing numeric bounds. upscale_images parameters upscaling_resize_w, upscaling_resize_h, and upscaling_resize lack min/max constraints. LLM could pass 1 or 100000, causing API errors.
Vague path handling. output_path parameters in generate_image and upscale_images lack validation rules: relative vs absolute paths? Existence checks? Directory creation? File overwrite behavior?