Automatic Exact-Tile Prompt + Cache

A ComfyUI node in Smart Upscaler/Prompting with 11 inputs and 5 outputs.

By HallettVisual·Created 21 days ago·Updated 12 days ago· 11

◄instruction—►

◄tile_reference—►

◄caption_generationManaged by Prompt Director (recommended)►

◄cache_moderead_write►

◄cache_tagqwen3vl_4b_fp8_local_target_v29_structured_detail_v13►

◄negative_prompt_fallbackchanged geometry, duplicated objects, artifacts, seams►

◄caption_max_side1344►

◄vision_model_idQwen3-VL-4B-FP8►

CategorySmart Upscaler/Prompting

Inputs (11)

Name	Type	Default	Description
clip	CLIP		—
image	IMAGE		—
instruction	STRING		—
tile_reference	STRING		—
prompt_system	SMART_PROMPT_SYSTEM		—
caption_generation	COMBO	Managed by Prompt Director (recommended)	Leave Managed. Tile detail and context now come from the Prompt Director; older choices remain only for loading older workflows.
cache_mode	COMBO	read_write	Reuse saved captions, regenerate them, or skip the cache.
cache_tag	STRING	qwen3vl_4b_fp8_local_target_v29_structured_detail_v13	Change this after changing the caption model or caption contract.
negative_prompt_fallback	STRING	changed geometry, duplicated objects, artifacts, seams	—
caption_max_sideopt	INT	13440–4096	Longest side of each real (unpadded) tile sent to the vision model. 1344 is the balanced 16 GB default; 0 sends the full tile, while 768-1024 saves more VRAM with a small fine-detail cost.
vision_model_idopt	STRING	Qwen3-VL-4B-FP8	Cache identity for the connected caption model. Change this when replacing or updating that model; this prevents old tile captions from another model being reused.