4B / 9B / Base vs Qwen 2511 Single-Image Comparison Workflow is a ComfyUI comparison workflow designed for testing single-image editing performance across multiple modern image models. It compares FLUX.2 Klein 9B, FLUX.2 Klein 4B, FLUX.2 Klein Base 4B, and Qwen Image Edit 2511 in one practical workflow, allowing creators to evaluate how different models handle the same source image and editing instruction.

This workflow focuses on single-image editing rather than multi-image reference transfer. The main purpose is to test how well each model can understand one input image, preserve important visual elements, and apply a targeted edit based on text instructions. It is especially useful for comparing background replacement, clothing changes, lighting enhancement, portrait refinement, object-preserving edits, and scene transformation while keeping the original subject identity and composition stable.

The FLUX.2 Klein 9B branch uses the flux-2-klein-9b-fp8 model with qwen_3_8b_fp8mixed text encoder and flux2-vae. This route is useful for testing stronger instruction following and better image understanding. Since the 9B model has more capacity than the 4B variant, it is suitable for edits that require precise preservation, such as keeping the subject’s face, pose, framing, hand position, and background layout while changing only one major element.

The FLUX.2 Klein 4B branch uses the flux-2-klein-4b-fp8 model with qwen_3_4b text encoder and flux2-vae. This route is useful for testing faster and lighter single-image editing. It may be easier to run online or on lower-resource setups, making it a practical choice for fast previews, quick testing, and efficient workflow deployment. Comparing it against 9B helps users judge whether the larger model’s quality gain is worth the extra resource cost.

The FLUX.2 Klein Base 4B branch provides another reference point for model behavior. It helps users compare the base version against the tuned 4B and 9B variants. This is useful for observing differences in conservativeness, edit strength, identity preservation, texture quality, and prompt adherence. The base model may preserve some structures differently from the tuned versions, so it is valuable for controlled model evaluation.

The Qwen Image Edit 2511 branch gives the workflow a strong editing-oriented comparison route. It is useful for testing instruction-based image editing against the FLUX.2 Klein family. Qwen Image Edit 2511 is often valuable for direct edit commands such as “replace the background,” “change the clothing,” “keep the subject unchanged,” or “add light and shadow to the image.” By placing Qwen 2511 in the same workflow, users can quickly compare whether Qwen or FLUX.2 Klein performs better for a specific single-image editing task.

A key part of this workflow is the ReferenceLatent structure. The source image is encoded into latent space through VAEEncode and used as a reference condition. This allows each model branch to receive the original image as structural guidance. The workflow is therefore not just generating a new image from text; it is editing from an existing image while attempting to preserve important parts of the original.

The workflow includes several practical editing scenarios. One prompt route tests background replacement, such as replacing an urban or indoor background with a quiet coastal cliff at an overcast sunset while preserving the subject’s pose and framing. Another route tests clothing replacement, such as changing a bathrobe into a fashionable modern outfit while preserving face, skin tone, pose, expression, hand position, chair, wall, plants, baskets, and other background elements. Another route tests portrait light and shadow enhancement, focusing on realistic skin texture, facial detail, natural light, and a clean portrait finish.

These tasks are useful because they reveal different model weaknesses. A model may generate a beautiful image but fail to preserve the face. Another model may keep the subject stable but not change the requested element strongly enough. Another may follow the edit but damage the hands, background, or original composition. This workflow makes those differences visible through direct side-by-side testing.

The workflow also includes model-specific output saving and image concatenation logic. The generated results can be saved separately by model branch, then combined into a comparison layout. This makes it useful for Civitai posts, YouTube model review videos, Bilibili workflow demonstrations, RunningHub showcases, and practical model selection before building a production editing pipeline.

Main features:

- Single-image editing comparison workflow

- FLUX.2 Klein 9B FP8 branch

- FLUX.2 Klein 4B FP8 branch

- FLUX.2 Klein Base 4B branch

- Qwen Image Edit 2511 branch

- ReferenceLatent-based source image preservation

- VAEEncode reference image conditioning

- Flux2Scheduler and SamplerCustomAdvanced support

- Same-image, same-instruction comparison

- Background replacement testing

- Clothing replacement testing

- Portrait lighting enhancement testing

- Subject identity preservation evaluation

- ImageConcanate side-by-side comparison output

- SaveImage output for each model branch

- Useful for Civitai model comparison and workflow publishing

Recommended use cases:

Single-image editing tests, model comparison, FLUX.2 Klein 9B testing, FLUX.2 Klein 4B testing, FLUX.2 Klein Base testing, Qwen Image Edit 2511 testing, portrait editing, background replacement, clothing replacement, lighting enhancement, subject-preserving edits, identity preservation testing, product-style edits, commercial image correction, AI model review, ComfyUI workflow evaluation, RunningHub online workflow publishing, and Civitai comparison posts.

Suggested workflow:

Start by uploading one clear source image. The source image should have a visible subject, stable composition, and enough detail for the model to preserve. Portraits, fashion shots, product images, seated character images, and clean social media images are good choices. Avoid extremely blurry images, heavily compressed images, very small subjects, or images where the object you want to preserve is unclear.

Then write a direct editing prompt. Since this is a single-image editing workflow, the prompt should clearly explain what should change and what should remain unchanged. For example: “Replace the background with a quiet coastal cliff at overcast sunset. Remove all buildings and streets. Add wind-shaped grass and a distant ocean horizon. Keep the subject’s pose and framing unchanged.” This type of prompt is better than a vague style prompt because it separates the edit target from the preserved elements.

For clothing edits, be very explicit about preservation. A good prompt should say what clothing should change and what must stay the same. For example, you can ask the model to change a bathrobe into a modern outfit while preserving the exact facial features, skin tone, pose, expression, hand position, chair, wall, plants, baskets, and other background elements. This helps reveal which model is better at local semantic editing without over-changing the whole image.

For lighting tests, describe the final lighting style rather than changing the entire identity of the image. You can ask for natural light, stronger facial modeling, realistic skin texture, visible pores, soft shadow, shallow depth of field, or high-quality portrait photography. This kind of test is useful for checking whether the model can enhance an image without making it look like a completely different person.

Keep the comparison controlled. Use the same source image, same prompt, same or similar resolution, and similar sampling settings when comparing model branches. The value of this workflow comes from controlled comparison. If every branch uses a different image or a different edit instruction, the comparison becomes less reliable.

When evaluating results, do not only judge which image looks the most beautiful. Look at whether the edit was actually completed. Check whether the subject identity is preserved, whether the face changed, whether the hands are damaged, whether the background was replaced correctly, whether the requested object or clothing was changed, and whether unwanted parts of the image were modified.

For background replacement, check whether the new scene is inserted cleanly while the subject remains unchanged. For clothing replacement, check whether the outfit changes without destroying body shape, hands, pose, chair, or background. For portrait lighting, check whether the image gains better shadow and skin detail without becoming plastic, over-smoothed, or overly artificial.

The 9B branch is useful when the edit is complex and preservation matters. The 4B branch is useful when speed and resource efficiency matter. The Base branch helps show the underlying behavior of the model family. The Qwen 2511 branch gives a strong image-editing baseline, especially for direct instruction-based edits.

This workflow is designed for creators who want to evaluate real editing ability rather than only text-to-image aesthetics. It helps test which model is better for background replacement, clothing editing, portrait refinement, identity preservation, and controlled single-image transformation. By comparing FLUX.2 Klein 9B, FLUX.2 Klein 4B, FLUX.2 Klein Base 4B, and Qwen Image Edit 2511 in one workflow, users can make more practical decisions before choosing a model for commercial image editing or content production.

🎥 YouTube Video Tutorial

Want to know what this workflow actually does and how to start fast?

This video explains what the tool is, how to launch the workflow instantly, and shares my core design logic — no local setup, no complicated environment.

Everything starts directly on RunningHub, so you can experience it in action first.

👉 YouTube Tutorial: https://youtu.be/7byGk8b04Ao

Before you begin, I recommend watching the video thoroughly — getting the full context helps you understand the tool faster and avoid common detours.

⚙️ RunningHub Workflow

Try the workflow online right now — no installation required.

👉 Workflow: https://www.runninghub.ai/post/2013130253551542274?inviteCode=rh-v1111

If the results meet your expectations, you can later deploy it locally for customization.

🎁 Fan Benefits: Register to get 1000 points + daily login 100 points — enjoy 4090 performance and 48 GB super power!

📺 Bilibili Updates (Mainland China & Asia-Pacific)

If you’re in the Asia-Pacific region, you can watch the video below to see the workflow demonstration and creative breakdown.

📺 Bilibili Video: https://www.bilibili.com/video/BV1EKkuBbE4R/

☕ Support Me on Ko-fi

If you find my content helpful and want to support future creations, you can buy me a coffee ☕.

Every bit of support helps me keep creating — just like a spark that can ignite a blazing flame.

👉 Ko-fi: https://ko-fi.com/aiksk

💼 Business Contact

For collaboration or inquiries, please contact aiksk95 on WeChat.

🎥 YouTube 视频教程

想了解这个工作流到底是怎样的工具，以及如何快速启动？

视频主要介绍 工具定位、快速启动方法 和 我的构筑思路。

我们会直接在 RunningHub 上进行演示，让你第一时间看到实际效果。

👉 YouTube 教程： https://youtu.be/7byGk8b04Ao

开始前建议尽量完整地观看视频 —— 把握整体思路会更快上手，也能少走常见弯路。

⚙️ 在线体验工作流

现在就可以在线体验，无需安装。

👉 工作流： https://www.runninghub.ai/post/2013130253551542274?inviteCode=rh-v1111

打开上方链接即可直接运行该工作流，实时查看生成效果。

如果觉得效果理想，你也可以在本地进行自定义部署。

🎁 粉丝福利： 注册即送 1000 积分，每日登录 100 积分，畅玩 4090 体验 48 G 超级性能！

📺 Bilibili 更新（中国大陆及南亚太地区）

如果你在中国大陆或南亚太地区，可以通过下方视频查看该工作流的实测效果与构思讲解。

📺 B站视频： https://www.bilibili.com/video/BV1EKkuBbE4R/

我会在 夸克网盘 持续更新模型资源：

👉 https://pan.quark.cn/s/20c6f6f8d87b

这些资源主要面向本地用户，方便进行创作与学习。