Flux Load References

Flux.2's signature move is multi-reference generation: give it up to ten images and it composes a consistent character, object, or style across them - no LoRA training required. In raw ComfyUI that means a small assembly line: a CLIPTextEncode, a LoadImage per reference, a VAEEncode each, and then the ReferenceLatent node chained once per image to fold each reference into the conditioning. Flux Load References (class DiztraidoReferenceChain) collapses that into one node: connect clip and vae, type a prompt, set reference_count, pick the reference images, and it hands you the finished positive conditioning plus a blank negative conditioning, ready for a sampler.

This is the node that makes the pack's Flux.2 story work end to end: load models with DiztraidoLoadFlux2Models, build references here, sample with the Flux Sampler.

How it works

Internally it runs ComfyUI's native nodes in sequence. It encodes your text_prompt with CLIPTextEncode, then for each active reference it loads the image (LoadImage), encodes it to latent space (VAEEncode), and appends it to the conditioning via ReferenceLatent - each reference chained after the last, in slot order. Finally it encodes an empty string to produce blank_negative, which is the correct negative for Flux.2 guidance-distilled workflows (you generally don't want a meaningful negative there). The references are the Flux.2 mechanism itself, not a ControlNet or a LoRA - you get the model's native multi-reference behavior.

The inputs that matter

• clip and vae - from your Flux.2 loader.
• text_prompt - the positive prompt, multiline.
• reference_count - 0 to 16 active reference slots. Set this first; it controls how many pickers matter.
• image_ref_1 … image_ref_16 - dropdowns of images in ComfyUI/input/. Slots beyond reference_count are ignored.
• Optional image_input_1 … image_input_16 - the same slots but as IMAGE sockets, so you can feed a reference from another node (an upscaled or preprocessed image) instead of a file on disk. If both are connected for a slot, the socket wins.

Outputs: positive (CONDITIONING) and blank_negative (CONDITIONING) - wire both into the sampler.

How to install

Part of Diztraido Nodes. ComfyUI Manager → search "Diztraido Nodes" → install diztraido-nodes → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/jadervasque/ComyUI-Diztraido.git

Restart and it's under the Diztraido flux category. No extra dependencies; it calls native ComfyUI nodes, so you need a ComfyUI build that ships ReferenceLatent (Flux.2-era ComfyUI).

Common issues

The two things people trip on: forgetting that reference_count gates everything - you can fill ten image pickers and see nothing happen because the count is still 0 - and the VAE/CLIP expectation, since this node needs the Flux.2 loader's clip and vae wired in before anything else. Also set expectations on reference strength: Flux.2 references are powerful, but a heavily-cropped or low-quality reference propagates its flaws into the output - feed it clean, well-lit images. And if you're on Klein rather than dev, reference behavior matches the model's distillation: same mechanism, but don't expect identical prompt adherence to dev. It's also not a ControlNet: for strict pose or layout control you'd still reach for other tools.