DEV Community

AI-KSK
AI-KSK

Posted on

Flux2 Klein Was Underrated: One Enhancement Node Turns Random Generation into Controlled Editing

#comfyui #ai #tutorial

Flux2 Klein processes text and reference images on separate paths and only merges them late in the denoise. The enhancement node exposes that split, which is why you can dial the prompt up while holding identity down — or trade identity for colour fidelity — without leaving the model. Six example graphs ship with the mirror.

Evidence note: this episode ships no workflow JSON, so no node counts or sampler values are claimed. The parameter values quoted below are the author's spoken settings, not values read out of a graph.

When this is useful

  • Holding a face steady while the pose changes.
  • Compositing two references where colour from one must not bleed into the other.
  • Cases where a single edit needs both a strong prompt and a strong reference, and you need to choose the balance per shot.
  • Upscaling via the sixth graph the author added to the five official examples.

How this differs from the previous build

The mental model matters more than any single number.

  • Two paths, late merge. Text goes one way, reference images another; they combine only in the second half of sampling. Control over that split is what the node sells.
  • Identity control comes in two flavours. A macro node pulls the result toward the reference latent (weight 0.5, start 0, end 0.8, bypassed by default); a micro node works inside attention layers on faces and hair (default 0.2, comfortable at 0.4, breaking above 0.5, with start_blocks 8 to skip the double blocks).
  • A balance node trades prompt against reference on a 0-1 dial, for the shots where you cannot decide up front.
  • Colour control is per-reference. In the dual-image graph, indices 0 and 1 let you set one reference's weight to 0 so the palette follows the other.

Running it

  1. Deploy the mirror, open ComfyUI, and pick a graph from the left panel — five official examples plus the upscale graph the author added.
  2. Load the image and write the prompt. Every one of the six starts the same way.
  3. Set size with the scale-image pixel node. Official guidance allows up to 4 (4M pixels); the author recommends staying at or below 2.
  4. To switch to the specialised model, change it at the UNet loader. The CLIP can be swapped for an uncensored variant.
  5. For identity work, prefer the attention-level node over the latent-pull node, and keep the weight near 0.2 before testing higher.
  6. For dual-image work, use the colour control node and index each reference.

Honest limits

  • The weights and block numbers are the author's tested settings, not shipped defaults read from a file — no JSON exists for this episode.
  • Above 0.5 on the attention node, results break in the author's own tests.
  • The macro identity node is bypassed by default and the author explicitly does not recommend it.
  • Flux2's own pixel ceiling of roughly 4M is above what the author advises; going past 2 trades quality for size.
  • The node is not widely known yet, which is exactly why a mirror exists — support is community-level.

Episode resources

Mirror (Compshare, referral code preserved)

Top comments (0)