RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2103123291263606786?inviteCode=rh-v1111
Qwen Image 2.1 — Visual-Prompt Reconstruction
Purpose
Generate an image-conditioned reconstruction from an empty latent, guided by a visual description and fixed preservation instructions. This is not a conventional pixel-preserving upscaler or a verified restoration benchmark.
Preparation and saved setup
Supply an authorized source image at node 514. The RHLLM node selects google/gemini-3.1-flash-lite-preview and sends its response into concat 545 alongside fixed text 544. This linked result overrides the encoder's old prompt. Review the service's routing, data terms, account requirements and response quality before use; no request was made in this review.
Encoder 512 receives the image, but its VAE port is unlinked and its latent output unused. Sampler 507 starts from EmptyLatentImage 506. Confirm compatibility with the installed encoder implementation instead of claiming a source-encoded low-denoise path.
The model uses the detail LoRA at approximately 0.5. Saved sampling controls are 25 steps, CFG 1, Euler/simple and denoise 1, with seed randomization. The preservation prompt is not an enforced identity, logo, text or geometry guarantee.
Size and output
Size controls use s=min(4,sqrt(2048²/(width*height))). This is an area budget and maximum×4 per-axis enlargement, not fixed 2048 dimensions. Large inputs can shrink; small inputs may remain below the budget; long sides can exceed2048. Integer conversion and latent-grid handling require output inspection.
Decoded output goes to one SaveImage and the reconstruction side of an original/output comparer. The source's 15 example groups do not constitute15 supplied workflows or new tests.
Assets
qwen_image_2.1_bf16.safetensorsqwen3vl_8b_fp8_scaled.safetensorsqwen_image_2.1_vae_bf16.safetensorselusarcas-qwen2-1-detailer-v1.safetensors
Files and version note
The historical comparer gallery is cleared in new copies; other data and originals are retained. This is local descriptive copy, not a published model upgrade. Covers are not verified results.
Tutorial and rights
Source-provided tutorial, not live-verified. The workflow is supplied through AI-KSK. No model run, image acceptance or performance measurement is included.
Input images and model weights are not bundled. Check workflow, model, LoRA, node, service, artwork, source-image and cover rights separately. Reconstruction does not grant ownership or a blanket commercial license.
Bilibili tutorial: https://www.bilibili.com/video/BV1SxaT6nEvk
Description
RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2103123291263606786?inviteCode=rh-v1111
Qwen Image 2.1 — Visual-Prompt Reconstruction PurposeGenerate an image-conditioned reconstruction from an empty latent, guided by a visual description and fixed preservation instructions. This is not a conventional pixel-preserving upscaler or a verified restoration benchmark.
Preparation and saved setupSupply an authorized source image at node 514. The RHLLM node selects google/gemini-3.1-flash-lite-preview and sends its response into concat 545 alongside fixed text 544. This linked result overrides the encoder's old prompt. Review the service's routing, data terms, account requirements and response quality before use; no request was made in this review.
Encoder 512 receives the image, but its VAE port is unlinked and its latent output unused. Sampler 507 starts from EmptyLatentImage 506. Confirm compatibility with the installed encoder implementation instead of claiming a source-encoded low-denoise path.
The model uses the detail LoRA at approximately 0.5. Saved sampling controls are 25 steps, CFG 1, Euler/simple and denoise 1, with seed randomization. The preservation prompt is not an enforced identity, logo, text or geometry guarantee.
Size and outputSize controls use s=min(4,sqrt(2048²/(width*height))). This is an area budget and maximum×4 per-axis enlargement, not fixed 2048 dimensions. Large inputs can shrink; small inputs may remain below the budget; long sides can exceed2048. Integer conversion and latent-grid handling require output inspection.
Decoded output goes to one SaveImage and the reconstruction side of an original/output comparer. The source's 15 example groups do not constitute15 supplied workflows or new tests.
Assetsqwen_image_2.1_bf16.safetensorsqwen3vl_8b_fp8_scaled.safetensorsqwen_image_2.1_vae_bf16.safetensorselusarcas-qwen2-1-detailer-v1.safetensors
The historical comparer gallery is cleared in new copies; other data and originals are retained. This is local descriptive copy, not a published model upgrade. Covers are not verified results.
Tutorial and rightsSource-provided tutorial, not live-verified. The workflow is supplied through AI-KSK. No model run, image acceptance or performance measurement is included.
Input images and model weights are not bundled. Check workflow, model, LoRA, node, service, artwork, source-image and cover rights separately. Reconstruction does not grant ownership or a blanket commercial license.
Bilibili tutorial: https://www.bilibili.com/video/BV1SxaT6nEvk