SquiDavey's ZIT Workflow (Local LLM / Styles / T2I / SeedVR2 / Post-Process) — v1.0

This ComfyUI workflow is for creating images from text concepts or upscaling existing images; it can expand prompts with a local LLM, generate with Z-Image Turbo, apply CRT/VHS processing, and perform tiled SeedVR upscaling, using a text concept or uploaded image as input and​​⁣⁣

For new images, LM Studio expands the concept into a richer prompt, Z-Image Turbo generates a base image, CRT/VHS processing adds analog effects, and SeedVR performs tiled upscaling.

For an existing image, upload it to the UPSCALE BYPASS node so the workflow uses it instead of starting with new image generation.

The text-to-image route takes a simple concept entered in the Prompt group; LM Studio expands it into a rich, photorealistic description before generation.

The image-to-upscale route takes an image uploaded to UPSCALE BYPASS and processes it with CRT/VHS simulation and tiled SeedVR upscaling.

Set up LM Studio with a downloaded model such as Qwen 2.5 7B, open the Developer/Server tab, and start the server on port 1234.

In ComfyUI/models/, place z_image_turbo_bf16.safetensors in models/diffusion_models/, ae.safetensors in models/vae/, qwen_3_4b.safetensors in models/text_encoders/, seedvr2_ema_7b_fp16.safetensors in models/checkpoints/, ema_vae_fp16.safetensors in models/vae/, and 4x-Clear​​⁣⁣

The listed model files are ae.safetensors, ema_vae_fp16.safetensors, qwen_3_4b.safetensors, seedvr2_ema_7b_fp16.safetensors, and z_image_turbo_bf16.safetensors.

The listed node packs are comfyui_essentials, comfyui_ttp_toolset, comfyui-crystools, comfyui-custom-scripts, comfyui-easy-use, comfyui-kjnodes, comfyui-lmstudio-image-to-text-node, comfyui-mxtoolkit, crt-nodes, rgthree-comfy, and seedvr2_videoupscaler.

For new generation, set the SEEDVR / LOAD IMAGE switch to True, enter a concept in the Prompt group, and press Queue Prompt.

For existing-image upscaling, set the switch to False, upload your image to UPSCALE BYPASS, and optionally disable the LLM group to save time.

If upscaling runs out of VRAM, try reducing the SeedVR tile size and increasing the blocks to swap.

Suggestion · not verified

Before using the SeedVR path, verify that ema_vae_fp16.safetensors and seedvr2_ema_7b_fp16.safetensors are the intended files for this workflow.

Suggestion · not verified

The listing does not specify a license for this workflow.