FLUX.3 Klein x Qwen-Image 2.1 in ComfyUI: Batch Pose and Outfit Changes with a Pose x Outfit Matrix (6 Workflows)

One character. Hundreds of poses. Over a thousand outfits. The interesting question is not whether a model can change clothes or change a pose once, but whether you can plan a whole batch of combinations without clicking through them one by one.

This AI-KSK episode wires two community visual galleries into ComfyUI and adds a custom pose x outfit matrix node, then ships six workflows across two editing models: FLUX.3 Klein and Qwen-Image 2.1.

What is actually inside

  • Pose gallery: 288 pose groups, 1,439 pose reference images.
  • Outfit gallery: 1,407 outfit references.
  • Custom nodes (open source, AI-KSK/KSK-Nodes): PoseGalleryBatch, CostumeGalleryBatch, AIKSK_PoseCostumeMatrix, plus model loaders for the two stacks.
  • Six workflows: pose-only batch, outfit-only batch, and pose x outfit matrix, for each of the two models.

The galleries are community open-source projects by the same author; the matrix node is AI-KSK's integration layer. Nothing here retrains a model or bakes a character into weights. The character reference is an input image, and every result still depends on the prompt, sampler, source quality and the model itself.

The six workflows

# Model Mode What it does
01 Qwen-Image 2.1 Pose Pick a character image, batch through the pose gallery
02 Qwen-Image 2.1 Outfit Pick a character image, batch through the outfit gallery
03 Qwen-Image 2.1 Matrix Cross pose x outfit selections into a queued batch
04 FLUX.3 Klein Pose Same pose batch, FLUX.3 Klein stack
05 FLUX.3 Klein Outfit Same outfit batch, FLUX.3 Klein stack
06 FLUX.3 Klein Matrix Pose x outfit matrix on FLUX.3 Klein

The Qwen stack loads qwen3vl_8b_int8_convrot as text encoder with qwen_image_2.1_vae_bf16; the FLUX.3 Klein stack loads flux3-image-klein-int8-convrot, qwen3vl_4b_int8_convrot, flux2-vae and the qwen3.5_9b_flux3_prompt_upsampler_int8_convrot prompt upsampler. Exact node counts and wiring are in the JSON files; the tables above are the intended usage, not a benchmark.

Why a matrix node

If you want five poses and five outfits for one character, doing it by hand is 25 runs. The matrix node takes the selected poses and the selected outfits, builds the full cross product, and queues the combinations. You can also hand-pick individual pairs when you do not want the full grid.

That is the whole trick: the galleries make references easy to browse, and the matrix node turns browsing into a batch plan.

Running it

  1. Deploy the AI-KSK mirror on UCloud CompShare: https://www.compshare.cn/images/fFwdv0x1vlhz?referral_code=A7fTQUYY9lCDFJwA6MMvoq
  2. Start the instance, open ComfyUI, and load one of the six workflows.
  3. Pick a character reference image, choose poses and/or outfits, then run a single job or the matrix batch.
  4. The custom nodes are also on GitHub: https://github.com/AI-KSK/KSK-Nodes
  5. Local all-in-one pack (Quark): https://pan.quark.cn/s/20e19307e8ef

Video walkthrough: YouTube https://youtu.be/KtYimd16ApM | Bilibili https://www.bilibili.com/video/BV1yC4Y6KEPF/

Honest limits

  • The galleries are third-party community assets; their licenses are separate from the workflows and from the model weights.
  • The mirror bundles the environment and the workflows, not third-party model weights beyond what the platform provides.
  • Results vary with prompt, reference image quality and sampler settings. Nothing here guarantees identity lock or production-ready output without review.
  • Counts (288 pose groups, 1,439 pose images, 1,407 outfits) describe the galleries shipped with this release.

Links