Upload 3 files

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+LICENSE.pdf filter=lfs diff=lfs merge=lfs -text

LICENSE.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b82a2805162bde714a4eb27b9063c4fc3345d08a30be055134a6160e5430ba74
+size 137711

NOTICE.txt

@@ -0,0 +1,5 @@
+Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing.
+
+This repository distributes a Derivative of the Krea 2 Turbo model, as defined in that agreement: a LoRA adapter produced by distilling Krea 2 Turbo to a 2-step schedule. The Krea Model has been modified. This adapter is not an official Krea product and is not endorsed by Krea.ai, Inc.
+
+Use of this adapter is subject to the Krea 2 Community License Agreement (LICENSE.pdf in this repository) and the Krea Acceptable Use Policy (https://krea.ai/krea-2-use-policy), exactly as use of Krea 2 Turbo itself is.

README.md

@@ -0,0 +1,191 @@
+---
+base_model: krea/Krea-2-Turbo
+base_model_relation: adapter
+license: other
+license_name: krea-2-community-license
+license_link: https://huggingface.co/lvladikov/Krea2-Turbo-Distill-2step-LoRA/blob/main/LICENSE.pdf
+library_name: diffusers
+tags:
+  - lora
+  - text-to-image
+  - distillation
+  - step-distillation
+  - distribution-matching
+  - krea-2
+pipeline_tag: text-to-image
+---
+
+# Krea 2 Turbo — 2-Step Distillation LoRA
+
+> 🚧 **Work in progress.** This page is the live account of the project — what the adapter is meant to do, how it is
+> being trained, and where it stands. It is updated as the recipe changes. There are no shipped files yet; the samples
+> and side-by-side comparisons will be added when there is a candidate worth showing.
+
+A LoRA for **[Krea 2 Turbo](https://huggingface.co/krea/Krea-2-Turbo)** that takes the model from its usual **8 steps
+down to 2** — Turbo's own weights and its own two sigmas, guidance 0.0, a quarter of the denoising passes — with the
+aim of matching what the **[4-step LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA)** delivers
+today, at every resolution it delivers it at.
+
+- 🎯 **The bar** — the [4-step LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA)'s output quality, across the same 12 resolutions, in two steps.
+- 🔌 **Drop-in, no exceptions** — a plain LoRA sampled by stock Euler at sigmas `[1.0, 0.5128]` in diffusers, ComfyUI
+  or MLX. No custom sampler, no policy head, no per-step tricks. If the quality needs a special sampler it is not this
+  project.
+- 🧬 **Same shape as the [4-step adapter](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA)** — rank 64 on the same 228 modules; a second adapter exists during training
+  only and never ships.
+- 📊 **Distribution matching, not imitation** — the training objective that finally moved this (see [Method](#method)).
+- 🎲 **The same 13,750 recorded teacher trajectories** the [4-step adapter](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) trained on, reused without a single teacher
+  re-run.
+- 🖥️ **One RTX 3090**, and a recipe shaped by its 24 GB.
+
+## Where it stands
+
+| | |
+| --- | --- |
+| lineage | [4-step LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) (78,000 samples) → 9,000 samples of 2-step trajectory distillation → distribution matching from there |
+| current run | **run 2**: distribution matching + a critic judged against the teacher's own finals, per-resolution weights measured from a full 12-bucket sweep |
+| best result so far | run 1 at 1,000 samples: at 512×512 and 768×1024 the sharpest, most coherent 2-step renders of the project — crowds resolved into people, faces intact at 768×1024; at 1280×1280 and above the same weights over-render (fine texture at 1.42× the teacher's), which run 2 addresses |
+| known gaps | faces at 512×512 still carry a faint doubled contour; fine structure at 1280×1280+ renders as fragments rather than the teacher's objects — the two things run 2 exists to fix |
+
+## How I got here
+
+The [4-step adapter](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) closed its page with a promise: a 2-step LoRA as the next project, and a guess at the lever it would
+need — matching the teacher's *distribution* rather than its trajectory. That guess turned out to be the whole story.
+
+The project began where the [4-step one](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) ended, from its final weights, and ran the same recipe at two steps:
+**progressive distillation** on the recorded teacher trajectories, each student call covering four teacher steps, with
+the LADD-style critic as the finisher. Nine thousand samples in, every number had stopped moving and the pictures had a
+signature the numbers could not see: doubled contours on faces and limbs, soft fine texture, crowds averaged into
+translucent overlaps. Several variations followed — the critic re-weighted, judged per token, a heavier hand on the
+final call, the student's own first-step output fed into its second — and each traded one of those faults for another
+without moving past them. A capacity probe ruled out adapter rank; a learning-rate shock ruled out the optimiser.
+
+The reason is structural, and worth stating plainly because it decides the whole design. A regression loss asks the
+student to land on the teacher's *specific* image for each prompt. When a two-step jump is wide enough that several
+images are plausible, the answer that minimises the squared error is their average — and the average of two sharp
+images is a blurred one with doubled edges. Every earlier recipe rewarded that average. Tuning its weights could not
+change what it rewarded.
+
+**Distribution matching** asks a different question: not "does your image match this one" but "would the teacher
+plausibly have produced your image". The first run of that objective, on top of the 9,000-sample weights, produced
+in a thousand samples what twenty thousand samples of the old recipe never had — and it did so while every latent
+distance to the teacher *rose*, which is exactly what a mode-seeking objective predicts and what a mean-seeking metric
+punishes. The distances are reported on this page; they are not optimised for, and they are not what decides a
+checkpoint. Pictures are, at fixed seeds, at every resolution, with faces viewed at 1:1.
+
+The recipe adjustments so far, each made on the measurement of the one before:
+
+1. progressive distillation at two steps from the [[4-step adapter](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA)'s](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) weights, with the LADD critic — the baseline
+2. the critic made to judge the first call's endpoint against the teacher's mid-states, which removed the gross
+   ghosting and left the doubled contours
+3. the critic's weight, its per-token form and the student's own first-step output as the second call's input —
+   tried one at a time; the objective flip that mattered was not among them
+4. **distribution matching** (the DMD2 family) as the primary objective, trajectory regression demoted to an anchor at
+   half weight, the critic switched off to read the new term alone
+5. the running average of the weights restarted at the objective switch, after it was caught averaging two lineages
+   into composites
+6. per-resolution weights for the distribution term, measured from a full 12-bucket sweep rather than estimated
+7. the critic back in, in its DMD2 form — on the trained score network's features, judged against the teacher's own
+   finals, conditional on the prompt, its gradient capped against the distribution term's
+
+## Method
+
+**Distribution matching with a trajectory anchor**, Krea 2 Turbo as its own teacher, on the recorded 8-step
+trajectories.
+
+The student makes two calls, at σ = 1.0 and σ = 0.5128 — the first and fifth points of the teacher's 8-step grid at
+mu = 1.15 — and stock Euler carries it between them. That grid is what makes the objective a drop-in: Euler's first step
+from pure noise lands *exactly* on the flow-matching interpolant at σ = 0.5128 with the same noise and the student's
+own clean-image prediction as the data point. So the student's first-call output is a legitimate image prediction that
+can be judged as an image, and the second call is fed from it during training the way it will be at inference.
+
+**The distribution term.** For an image the student produces, two denoisers estimate how it should be cleaned up from a
+freshly noised copy: the frozen teacher, and a second small adapter on the same frozen base — the *fake score* — that
+is trained online to denoise whatever the student currently makes. Where the two disagree is the direction that makes
+the image more like the teacher's work and less like the student's habits, and the student is pushed that way
+(the DMD2 gradient, per-sample normalised). Averaging is never rewarded, so the student commits. The fake adapter is
+rank 32, starts as an exact copy of the teacher, updates twice per student step, and is discarded at the end.
+
+**The anchor.** Plain trajectory regression on the teacher's recorded chords stays in at half weight. It keeps the
+student on the teacher's two-step grid so the distribution term cannot wander into a different sampler behaviour, and
+it is what the 9,000 earlier samples had already satisfied — which is why the first distribution-matching run moved
+so far so fast.
+
+**Per resolution.** The distribution term's push grows with resolution: the fake adapter sees few large-bucket samples
+and under-fits fine structure there, and a per-pixel normaliser lands harder as pixel counts grow. Rather than guess a
+taper, a full 12-bucket sweep of the run-1 weights measured the fine-texture ratio to the teacher at every resolution,
+and the term's weight per bucket is set from that measurement so that each bucket is pushed toward the teacher's
+texture rather than past it.
+
+### The critic (run 2)
+
+The fake adapter cannot notice its own mistakes, and at large resolutions those mistakes render as fragments —
+right amount of detail, wrong structure. A **discriminator** can: trained to tell the student's images from the
+teacher's, it learns whatever systematic difference exists between them, fragments included, and its gradient points
+back at the teacher. This is the second half of DMD2 and it differs from the [[4-step adapter](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA)'s](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) critic in two ways that
+follow from what the earlier runs taught:
+
+- it sits on the **trained fake adapter's features**, not the frozen teacher's, so it can learn a decision rather than
+  reweight what the teacher already computes;
+- its real class is the **teacher's own finals**, conditional on the prompt — the bar is the teacher — and its gradient
+  is **capped per sample** at twice the distribution term's, so it can correct but never take over.
+
+## What the LoRA touches
+
+Rank **64**, alpha = rank, bf16, the same **228 modules** as the [4-step adapter](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA): the 8 attention and feed-forward linears
+of all 28 transformer blocks, plus the four global linears — `time_embed.linear_1`, `time_embed.linear_2`,
+`time_mod_proj`, `final_layer.linear` — that a step-count change needs most. Nothing about the base model changes.
+
+## Training data
+
+The **13,750 recorded teacher trajectories** of the [4-step project](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) — Krea 2 Turbo's own 8-step run at mu = 1.15 and
+guidance 0.0, every latent and velocity stored — serve unchanged: a 2-step chord is two of the 4-step chords end to
+end. 203 held-out prompts measure the student–teacher gap on unseen prompts and never receive a gradient. The critic's
+real class is the teacher's finals for the training prompts; the 43,044 real-photo crops of the [4-step project](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) are
+available to it as a second source and are not used in the current run.
+
+## Resolutions
+
+The same 12 buckets as the [4-step adapter](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA), interleaved in proportion to their remaining samples:
+
+|           |           |           |
+| --------- | --------- | --------- |
+| 512×512   | 512×768   | 768×512   |
+| 768×768   | 768×1024  | 1024×768  |
+| 1024×1024 | 960×1280  | 1280×960  |
+| 1280×1280 | 1440×1280 | 1440×1440 |
+
+## Hardware
+
+One **RTX 3090 (24 GB)**. The frozen base is weight-only int8; the student's checkpointed block inputs stage to pinned
+host memory above 0.3 megapixels; the student, the fake adapter and the critic each build and free their own graph
+in turn, so their peaks never overlap; a hard memory ceiling sits below the driver's paging threshold so a step that
+does not fit fails loudly. A full step with every term live reserves about 21.5 GB at 1440×1440. The price of the
+objective is throughput: about 160 training samples an hour against the [4-step recipe](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA)'s 470 — three times the cost
+per sample, and so far a small fraction of the samples.
+
+## How it is judged
+
+Every 1,000 samples, both the live weights and their running average are pulled, merged and rendered at fixed seeds on
+15 fixed prompts across four resolutions, with faces cut out at 1:1, a fine-texture ratio and frequency bands against
+the teacher, a graded judge, and a pairwise preference against the teacher. Latent distances — the held-out chord gap
+and the two-step rollout error — are recorded but never used to keep or stop a run: this project's clearest lesson is
+that they reward blur, and a run that improved them while its pictures collapsed was stopped by its pictures.
+
+## Notes and limitations
+
+- 🎯 **Krea 2 Turbo only**, at **2 steps**, guidance **0.0** (cfg 1.0 in ComfyUI), **mu = 1.15** — the two training
+  sigmas are anchored to that grid.
+- 🚧 Not released. Everything on this page is the current state of a run in progress.
+
+## What's next
+
+Run 2 is training. Its first probe decides whether the critic removes the large-resolution fragments; if it does not,
+the next lever is a perceptual-space term on decoded image crops, which judges structure rather than energy. The
+shipping point, when there is one, is chosen by the sweep and the eye, not by distance run — the same discipline as the
+[4-step adapter](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA).
+
+## License
+
+The adapter is a derivative of [Krea 2 Turbo](https://huggingface.co/krea/Krea-2-Turbo) and is covered by the
+**Krea 2 Community License Agreement** (`LICENSE.pdf` in this repository), as the [4-step adapter](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) is. The attribution
+notice the license requires of a derivative ships with the files.