Distilled vs base checkpoints: Turbo or Raw for a face

This week the two most downloaded open image models for a recurring face came as pairs: Qwen-Image 2.1 and its 8-step Turbo, Krea 2 Raw and its 8-step Turbo. Same weights lineage, same license, very different behaviour. Krea's Raw runs 52 steps in Comfy's workflow; its Turbo runs 8 with guidance at 0.0. If you run a character, picking the wrong half of the pair costs you either hours or the likeness.

This explainer is the concept behind those file names: what distillation changes, what it quietly takes away, which setting breaks when you swap one for the other, and which half to train on. Every number below is read from a model card or a paper linked where it is used. For the model-specific side, see our Krea 2 vs Qwen-Image 2.1 comparison and the MiniMax H3 Turbo breakdown, which is the same idea applied to video.

What distillation is, in one paragraph

A base model removes noise from an image over many small steps, and usually runs twice per step: once with your prompt and once without, then pushes the result away from the blank one. That push is classifier-free guidance, the CFG number. Distillation trains a second copy, the student, to reach the same final image in a handful of steps and without the second pass. The student is what ships as Turbo or Lightning.

The pairs you will actually meet

What you gain

Speed first. Eight steps against forty is five times fewer steps, and if the base is run with guidance on, each of those forty costs two passes, so the arithmetic ceiling is roughly ten times less compute per image (our arithmetic, not a measured benchmark; real gains depend on the card, the text encoder and the VAE). Turbo drafts also let you reject a bad pose in seconds instead of minutes, which matters more for a daily feed than any quality score.

Steadiness second. Third-party Krea 2 reviews (not Krea's own claims) describe Raw as more varied and malleable with finer micro-texture, and Turbo as more consistent, polished and slightly smoothed. For a character, polish and consistency are often what you want in the final frame, as long as the face holds.

What you lose, and why

The loss is variety. Gandikota and Bau (arXiv 2503.10637) looked at how images form inside distilled and base models and report that distilled models settle their final structure almost immediately, at the first timestep, while base models spread those decisions across many steps. If the layout is fixed in step one, changing the seed changes little. Their fix, run the base for the first step and hand over to the distilled model, restored diversity in their tests.

The creator community found the same thing without the paper. Krea 2 users complained that Turbo outputs look alike across seeds, and a diversity adapter for Krea 2 Turbo now exists. Its author reports a rank-32 LoRA that raises pairwise CLIP distance from 0.0296 to 0.0548 (+85%) on a panel of 10 prompts by 16 seeds, with an image-quality score at parity. Read that as the author's own small test: a distance metric measures variation, not whether the extra variety is better.

docs.comfy.orgOpen the page

The settings that break when you swap

For a recurring face: train on the base, render on the distilled copy

Krea's own repository recommends training a LoRA on Raw and applying it to Turbo, and Comfy's page says the same. The logic: Raw is undistilled and malleable while Turbo is built for fast generation. That is the pattern to copy for any pair: teach the identity to the model that can still change, then render on the one that is cheap. Our consistent character guide covers the reference-set side of this.

There is one honest unknown. Neither lab publishes a face-similarity test for Turbo against the base, so "the likeness survives distillation" is a claim you verify on your own character: render the same 8 prompts on both with the LoRA at the same strength, and compare faces side by side before you move the feed to Turbo.

Gaps and caveats worth knowing

• A small blind test by a commercial API vendor rated Qwen Turbo's text-to-image quality below the base on 14 questions, while its edits held up. One vendor, small panel: a hint, not a verdict.
• A GGUF publisher reports Turbo drifting from the original picture faster than expected at 4 bits and below, so quantizing a distilled model compounds the loss.
• The diversity adapter is a research artifact: its author says it applies only to the first 2 of 8 steps, collapses quality if left on for all 8, and stock ComfyUI applies LoRAs on every step, so it needs a step-scheduling node.
• Licenses carry over. Qwen's Turbo is under the same Qwen Research License as the base and Krea's two checkpoints share the gated Krea 2 Community License, so distilling does not get you out of a commercial-use limit. Read the license before you monetize.

Our verdict

Treat Turbo as the production tool and the base as the teacher. Train and test on the base, render the daily feed on the distilled copy, and keep a base-model recipe for the shots that need a different composition. Do the side-by-side face test once per character, then trust it.

If you would rather not manage this by hand, the Musetta studio runs on your own ComfyUI and keeps each character's reference photos, so a draft on a fast checkpoint and a final on the base start from the same approved face. The advice above works in plain ComfyUI; how Musetta compares with working in ComfyUI directly is a separate page.

Musetta studioDemo data