RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2089343535781859329?inviteCode=rh-v1111

Anima 2.9B Smart Image-to-Image with Qwen3-VL Caption

Purpose: take an input image, let Qwen3-VL write an Anima-compatible prompt from it, and run a low-denoise image-to-image pass with the Anima 2.9B checkpoint. The point is to re-render an existing illustration in the Anima style without hand-writing the prompt.

Graph (as saved)

14 nodes / 16 links.

• UNETLoader Anima-2.9B-preview-v1.safetensors; CLIPLoader qwen_3_06b_base.safetensors (type stable_diffusion); VAELoaderKJ qwen_image_vae.safetensors.
• Input: LoadImage, then image_scale_pixel_v2 at ~1.05 MP (64 multiple) for the img2img encode.
• Caption: a Qwen3VLProcessor node carries the instruction to analyse the input image and output one Anima-compatible prompt (quality tags, year tags, character count, then subject and scene description).
• Sampling: VAEEncode, KSampler 28 steps / CFG 5 / res_multistep / linear_quadratic / denoise 0.7; a PreviewImage shows the result; SaveImage prefix Anima_2.9B/img2img.
• The saved CLIPTextEncode values are the tag "dog" and the official negative block — demo content, not the effective caption path.

How to use

Replace the input image and inspect the Qwen3-VL caption before queueing; the denoise (0.7) is high enough that a wrong caption changes the result more than the input does. Lower the denoise if you need the original composition preserved more strictly.

Limits and not-verified items

No inference was run for this listing: no caption accuracy, speed, VRAM or style-fidelity claim is made; values are quoted from the saved nodes. The VLM and its weights are not bundled and have their own license. The cover image is a promotional still, not a frame produced by this graph.

Tutorial: YouTube · Bilibili. Workflows supplied through AI-KSK; verify workflow, node, model, VLM, image and cover rights separately. No blanket commercial or redistribution license is asserted.