Qwen-Image 2.1 Fun ControlNet Union: 8 Controls in One File

Qwen-Image 2.1 landed on September 20, 2026 as a 7B single-stream DiT that generates and edits with the same weights and writes real RGBA alpha. The ControlNet-Union branch follows the same path Alibaba PAI took for MiniMax H3: a single control checkpoint that covers every condition, so no per-condition weight switching is needed.

One branch, eight conditions

The released file holds only the control branch (control_img_in plus 16 control_blocks, about 7.0 GB per the model card, 7.55 GB on disk). It is loaded with strict=False on top of the base Qwen-Image 2.1 transformer, which stays frozen, so the base weights are reused exactly as they are.

control_context_scale scales every control skip before it is added to the main branch: 1.0 is the strongest setting used for all published results, lower values weaken the guidance, and 0.0 switches the control branch off.

Control conditions

Any ordinary RGB control image at the target canvas works: the model card notes it tolerates different line thickness, thresholds and crops. All samples below come from the model card at 40 inference steps, control_context_scale = 1.0 and seed 43, with the control signal on the left and the generation on the right.

Canny edge map	Generated output

Depth map	Generated output

Grayscale image	Generated output

Scribble sketch	Generated output

The remaining conditions (HED, Lineart, MLSD, Pose) follow the same pattern on the model page.

Inpainting shares the same branch

The control input is widened to 129 channels precisely so one branch can do both jobs: control latents, a keep-mask and masked-image latents are concatenated before they enter the branch. For pure control the mask and masked-image channels are zero-padded, so the same checkpoint still runs plain Canny or Depth. For inpainting the masked region is re-drawn from the prompt while the rest of the frame is preserved, and the two can be combined: feed a control image together with a mask and the re-drawn region follows both the prompt and the given structure.

The mask is white where content should be regenerated and black where it should be kept. Regenerated pixels are encoded at mid-gray, the zero of the VAE's [-1, 1] input range, so untouched areas survive the round trip.

Source image	Mask	Pose control	Inpaint output

ComfyUI support

ComfyUI support is in review rather than released. Kijai opened PR #16519 ("Support Qwen-Image 2.1 union fun controlnet") on September 24, 2026, and it changes four files:

• comfy/ldm/qwen_image21/model.py adds QwenImage21FunControl and QwenImage21FunControlBlock, a VACE-style branch that seeds the control latents into the target rows and adds each block's after_proj skip back after one base block.
• comfy_extras/nodes_model_patch.py teaches ModelPatchLoader the layout: it detects the checkpoint by control_img_in plus control_blocks.0.img_mlp.out.weight and derives block count, control_in_dim, head dim and MLP ratio from the state dict, so one loader fits this checkpoint family. Quantized files load with mixed-precision ops, keeping the layers quantized with bf16 compute.
• comfy_extras/nodes_qwen.py adds QwenImage21FunControlNetApply ("Apply Qwen Image 2.1 Fun ControlNet"): inputs are model plus model_patch, strength (default 1.0), start_percent / end_percent, and optional control_image, inpaint_image and mask ("1 marks the region to regenerate").
• comfy/controlnet.py tightens the Qwen Fun loader detection so the new checkpoint is not mistaken for an older ControlNet layout.

Converted ComfyUI patches are already published by Kijai while the PR is in review:

• qwen_image_2.1_fun_controlnet_union_bf16.safetensors
• qwen_image_2.1_fun_controlnet_union_int8_convrot.safetensors

There is no official workflow template for it in Comfy-Org/workflow_templates yet, so the previous Qwen-Image templates (image_qwen_image_controlnet_patch, image_qwen_image_instantx_inpainting_controlnet) are for older models and do not apply here.

Availability

Outside ComfyUI the checkpoint runs through VideoX-Fun: clone VideoX-Fun, put the base Qwen-Image 2.1 model and the control checkpoint under models/Diffusion_Transformer/, and run examples/qwenimage21_fun/predict_t2i_control.py (or predict_i2i_inpaint.py for inpainting). The branch was added there in the commit "Update Qwen Image 2.1 Control and Flex Forcing".

Four notes from the model card are worth keeping:

• config_path must be config/qwenimage21/qwenimage21_control.yaml. It builds the control branch exactly as the checkpoint expects (control_layers: [0, 2, 4, ..., 30], control_in_dim: 129), and any other config silently drops or misplaces control weights and produces wrong output.
• sample_size sets the output canvas. Keep both sides as multiples of 16 so the control map is not distorted.
• use_kv_cache = True caches text and condition-image keys after the first denoising step, which speeds up generation at a fixed resolution.
• Memory: the transformer plus the Qwen3-VL text encoder do not fit a single consumer GPU fully loaded. The model card recommends model_group_offload as the fastest option, or model_cpu_offload_and_qfloat8 on a single high-memory GPU.