Flux2 Kontext Conditioner

If you've stared at a Flux 2 Klein editing workflow and wondered why "change the shirt" needs four nodes wired in exactly the right order, this is the node that gets rid of the wiring. Flux2 Kontext Conditioner is a one-box replacement for the whole Scale Image to Pixels → VAE Encode → Reference Latent → text encode chain: you feed it a CLIP, a VAE, and up to three reference images, and it hands you a ready conditioning plus the latent to sample from. Klein's Kontext attention is genuinely good at instruction edits - swap a garment, change the lighting, keep the person - but it only behaves when the reference latents are shaped right. This node encodes them correctly so you don't have to.

How it actually works

Under the hood it's the standard Flux 2 recipe, done carefully. Each connected image gets resized to its own imageN_resolution (scaled by area, not squashed), snapped to a multiple of alignment - 16 is the default and the minimum safe value, because the Flux2 VAE compresses spatially 16x - then VAE-encoded into the 16-channel latent the model expects. Each encoded latent is scaled by its strength and appended to the conditioning under the reference_latents key, which is exactly what Klein's attention reads. Your prompt runs through the Qwen3-8B text encoder (clip.tokenize → encode_from_tokens_scheduled, no vision tokens), and the guidance scale is baked straight into the conditioning, which is how CFG-distilled Flux 2 expects to be steered.

One thing the README buries: image1 is special. The latent output is image1's encoded latent, so image1 is both a reference and the starting canvas the sampler denoises. Images 2 and 3 are pure references. If you want to edit from a specific source image, that one goes in slot 1. Connect nothing and no VAE, and the node switches to text-to-image: an empty canvas latent at target_resolution (1024 default) plus plain text conditioning.

The inputs that matter

• image1 / image2 / image3 + strength + resolution - per-slot control. strength 1.0 is full influence; 0–2 range, above 1 amplifies for stronger style transfer (and artifacts - you've been warned). resolution is the VAE encoding res, default 1024, max 2048.
• guidance - default 3.5, baked in. Raise it for more prompt adherence, lower for naturalness. Leave it alone until you've got a feel for it.
• vae_tiling - flip this on at 1536px+ encoding res to cut peak VRAM at the cost of slower encodes.
• system_prompt_mode - an honest no-op-ish. Flux 2 doesn't natively process system prompts, and the author says the Qwen3 chat-template wrapping only subtly shifts token distribution. Set to none unless you're experimenting.

Outputs: conditioning and latent both go to the KSampler; VAE-decode the result. The three output_image1/2/3 are debug only - wire one to a Preview Image if you want to see what the resize/alignment snapping did. They're not needed for generation.

Installing

ComfyUI Manager, search "ComfyUI-Flux2KontextConditioner", or:

cd ComfyUI/custom_nodes
git clone https://github.com/KursatAs/ComfyUI-Flux2KontextConditioner

Then restart. That's the whole dependency story - it only needs torch, which you already have. The real work is the models: flux-2-klein-9b.safetensors, qwen_3_8b_.safetensors, and flux2-vae.safetensors. Grab all three before you blame the node. Worth knowing: the 9B Klein is BFL's non-commercial tier, same license politics that drove the community to Apache-2.0 editors like Qwen-Image-Edit; if that matters to you, its Apache-2.0 Klein 4B sibling exists.

Where people get burned

The big one is in the README's warning: don't pre-process your images. No Scale Image to Pixels, no VAE Encode, no Reference Latent ahead of this node - plug the raw image straight into image1. Feeding it an already-encoded latent double-encodes and produces garbage. Also, the whole point of this node is keeping the reference-latent plumbing self-contained; hand-rolling conditioning with reference_latents on another node is where people hit cuda/cpu tensor mismatch errors. And remember the KSampler has no decode built in - forgetting the VAE Decode on the other end is a classic 5-minute "why is this blank."