RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2107777471256297473?inviteCode=rh-v1111
Bilibili tutorial video: https://www.bilibili.com/video/BV1AsHy6nEPR/
# Qwen Image 2.1 — Product Harmonization, 40 and 20+6 Routes
## Purpose
Compare two configured harmonization routes using one active prepared composition, not an automatically assembled set of product/background inputs. Intended for users able to resolve compatible ComfyUI dependencies and review real outputs. This is a workflow resource, not a tested model-weight release.
## Asset references
`qwen_image_2.1_bf16.safetensors`, `qwen_image_2.1_int8_convrot.safetensors`, `qwen3vl_8b_int8_convrot.safetensors`, `qwen_image_2.1_vae_bf16.safetensors`, `Qwenimag21_c2-st2000.safetensors`, `Qwen-Image-2.1-viggle-turbo-v0.2.1-6step-lora-r256.safetensors`, `4x-UltraSharpV2.safetensors`.
## Saved configuration
32 nodes/40 links. LoadImage1210 goes through UltraSharp upscale895 and a roughly4MP resize896 before encoder494 image1. The legacy canvas chain is bypassed and ends at preview668; re-enabling it alone does not redirect generation. Prompt626 is linked into the encoder.
## How to prepare
Supply an authorized prepared composition and inspect its prompt and conditioning. Standard route: BF16-named base609 → C2 LoRA613 → sampler618 → decode490. Alternate: INT8-named base1200 → C2 LoRA1201 → sampler1202, then a Viggle LoRA branch and sampler1205 taking1202 latent. FluxGuidance1204 stores6 for the second positive-conditioning path, not sampler CFG6.
## Known limits
The author labels the alternate20+6, not six steps end-to-end. Saved arrays use26 and boundary20; inspect noise addition, seed control, and leftover-noise handling before assuming seamless continuation. Base filenames, LoRA processing, guidance and sampling all differ, so this is not a step-only benchmark. Comparer630 sees the two decoded outputs, not the original. Two savers per route are configured; actual files are unverified.
No new inference, visual/audio acceptance, benchmark, memory measurement, node-version certification, or license verification is included. Model filenames are asset references, not proofs of internal implementation or compatibility. Weights and selected input media are not bundled.
## Files and version note
- 02: [Workflow JSON]() · [Selected cover]()
Historical comparer-gallery data, where present, is cleared in the distribution copy; other data and the source original are retained. Selected covers are not verified inference examples. This local descriptive copy is not a publication announcement or a model upgrade.
## Tutorial, attribution, and rights
[Source-provided tutorial](https://www.youtube.com/watch?v=K6HI-iuMXeg), not live-verified. The workflows are supplied through AI-KSK; verify workflow, model, node, product-image, cover and audio rights separately. No universal commercial permission is asserted.
Description
RunningHub workflow page: https://www.runninghub.ai/zh-cn/post/2107777471256297473?inviteCode=rh-v1111
Bilibili tutorial video: https://www.bilibili.com/video/BV1AsHy6nEPR/
# Qwen Image 2.1 — Product Harmonization, 40 and 20+6 Routes
## Purpose
Compare two configured harmonization routes using one active prepared composition, not an automatically assembled set of product/background inputs. Intended for users able to resolve compatible ComfyUI dependencies and review real outputs. This is a workflow resource, not a tested model-weight release.
## Asset references
`qwen_image_2.1_bf16.safetensors`, `qwen_image_2.1_int8_convrot.safetensors`, `qwen3vl_8b_int8_convrot.safetensors`, `qwen_image_2.1_vae_bf16.safetensors`, `Qwenimag21_c2-st2000.safetensors`, `Qwen-Image-2.1-viggle-turbo-v0.2.1-6step-lora-r256.safetensors`, `4x-UltraSharpV2.safetensors`.
## Saved configuration
32 nodes/40 links. LoadImage1210 goes through UltraSharp upscale895 and a roughly4MP resize896 before encoder494 image1. The legacy canvas chain is bypassed and ends at preview668; re-enabling it alone does not redirect generation. Prompt626 is linked into the encoder.
## How to prepare
Supply an authorized prepared composition and inspect its prompt and conditioning. Standard route: BF16-named base609 → C2 LoRA613 → sampler618 → decode490. Alternate: INT8-named base1200 → C2 LoRA1201 → sampler1202, then a Viggle LoRA branch and sampler1205 taking1202 latent. FluxGuidance1204 stores6 for the second positive-conditioning path, not sampler CFG6.
## Known limits
The author labels the alternate20+6, not six steps end-to-end. Saved arrays use26 and boundary20; inspect noise addition, seed control, and leftover-noise handling before assuming seamless continuation. Base filenames, LoRA processing, guidance and sampling all differ, so this is not a step-only benchmark. Comparer630 sees the two decoded outputs, not the original. Two savers per route are configured; actual files are unverified.
No new inference, visual/audio acceptance, benchmark, memory measurement, node-version certification, or license verification is included. Model filenames are asset references, not proofs of internal implementation or compatibility. Weights and selected input media are not bundled.
## Files and version note
- 02: [Workflow JSON]() · [Selected cover]()
Historical comparer-gallery data, where present, is cleared in the distribution copy; other data and the source original are retained. Selected covers are not verified inference examples. This local descriptive copy is not a publication announcement or a model upgrade.
## Tutorial, attribution, and rights
[Source-provided tutorial](https://www.youtube.com/watch?v=K6HI-iuMXeg), not live-verified. The workflows are supplied through AI-KSK; verify workflow, model, node, product-image, cover and audio rights separately. No universal commercial permission is asserted.