chk17464 release

README.md

@@ -17,11 +17,22 @@ pipeline_tag: text-to-image
 
 # Krea 2 Turbo — 2-Step Distillation LoRA
 
-> 🧪 **A fast-preview adapter, from a project still in training.** The published checkpoint files give usable two-step
-> renders and are measured honestly below 4-step or 8-step renders; they are not the
-> [4-step LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA)'s quality, and that adapter remains the recommendation for quality renders. Training
-> continues one recipe change at a time, and a later checkpoint replaces this one only when the sweeps and I visually
-> agree it is better. Known issues — see [Known issues](#known-issues).
+> 🧪 **A fast-preview adapter, from a project still in training.** When the subject is close and fills a good part of the
+> frame — a portrait, a single figure, an object seen up close — two steps already hold up well, and you can rely on
+> this adapter for those images. Small subjects are where it still falls short, people and objects alike: faces in a
+> crowd, figures in a wide scene, the machines at the back of a gym — anything that takes up little of the frame can
+> come out ghosted or smeared. For those, and whenever quality matters more than speed, use the
+> [4-step LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA). Every figure on this page measures the
+> adapter honestly against 4-step and 8-step renders. Training continues one recipe change at a time, and a later
+> checkpoint replaces this one only when the sweeps and I visually agree it is better. Known issues — see
+> [Known issues](#known-issues).
+>
+> 📐 **The saved steps can also go into resolution.** A small subject is simply one that covers few pixels, so a larger
+> render makes the same subject bigger — and at a quarter of the teacher's steps, renders up to 2048×2048, Krea's
+> published maximum recommended resolution and beyond the largest size this adapter was trained at (1440×1440), come
+> within easy reach. That makes the adapter a stepping stone to high-resolution renders as well as a fast preview. Past
+> 2048×2048, stock Krea 2 itself begins to duplicate subjects — a property of the base model, with or without this
+> adapter.
 
 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 — aiming at
@@ -32,9 +43,9 @@ recommendation for quality renders.
 - 🎯 **The aim** — the best two-step quality this base can give, at every one of the same 12 resolutions, measured
   against the 8-step teacher and against the [4-step LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) as the reference. Not a claim to reach either.
 - ⚡ **A quarter of the steps** — 8 → 2, on Turbo's own deployment sigmas.
-- ⏱️ **3.8× faster denoising** — the model runs twice instead of eight times, and denoising is the part this adapter
-  changes: **79.9 s → 20.9 s** measured at 1024×1024 on the same prompts, the adapter itself costing about 3.5% per
-  call. What a whole render costs on top of that is unchanged by the LoRA and depends on your pipeline; see
+- ⏱️ **4.2× faster denoising** — the model runs twice instead of eight times, and denoising is the part this adapter
+  changes: **81.4 s → 19.5 s** measured at 1024×1024 on the same prompts, the adapter's own cost per call within
+  measurement noise. What a whole render costs on top of that is unchanged by the LoRA and depends on your pipeline; see
   [Performance](#performance).
 - 📊 **Distribution matching, not imitation** — the training objective that got the renders improving again after the
   4-step project's recipe had stopped helping at two steps (see [Method](#method)).
@@ -42,21 +53,23 @@ recommendation for quality renders.
   evaluated on each prompt's own conditioning, so the student is matched to what the teacher makes _for that prompt_,
   not to a prompt-free look. There is no separate adherence term: instead a vision-language judge checks every
   checkpoint — each render scored alone against the prompt's objects, counts, attributes and relations, with the
-  teacher scored the same way — and a term would only be added if that meter showed adherence slipping.
+  teacher scored the same way — and a term would only be added if that meter showed adherence slipping. The one part of
+  training that looks at images without their prompt is the artefact critic (see [Method](#method)), and it only judges
+  whether fine structure looks like the teacher's.
 - 📐 **12 trained resolutions** — multi-aspect from 512×512 up to 1440×1440, each with its sweep.
-- 🔌 **Drop-in, no exceptions** — a plain LoRA sampled by stock Euler at sigmas `[1.0, 0.5128]` in diffusers, ComfyUI
+- 🔌 **Drop-in, no exceptions** — a plain LoRA sampled by stock Euler at sigmas `[1.0, 0.7595]` 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.
 - 🎲 **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.
-- 🔢 **13,663 training samples** in the 2-step stages, on top of the [4-step LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA)'s 78,000 — all of them
+- 🔢 **17,464 training samples** in the 2-step stages, on top of the [4-step LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA)'s 78,000 — all of them
   drawn from the **same recorded material**: no new prompts, no new text embeddings and not one new teacher run. A
   training sample is one pass over a prompt that was already encoded and already traced by the teacher for the
   4-step project, read again at the two sigmas this schedule uses.
-- 📅 **5 days** from the first 2-step training launch to this checkpoint, on a single RTX 3090 — and the project continues.
-- 🔁 **15 recipe adjustments** across two methods so far — seven of trajectory distillation before the switch, eight of distribution matching since.
+- 📅 **7 days** from the first 2-step training launch to this checkpoint, on a single RTX 3090 — and the project continues.
+- 🔁 **26 recipe adjustments** across two methods so far — seven of trajectory distillation before the switch, nineteen of distribution matching since — each kept only when the renders did not get worse.
 - 🖥️ **One RTX 3090**, and a recipe shaped by its 24 GB.
 
 [![The 15 test prompts, rendered by Krea 2 Turbo with this LoRA at 2 steps](assets/thumbs/poster.jpg)](assets/poster.jpg)
@@ -86,13 +99,13 @@ place to check which checkpoint the current files are based on.
 
 |               |                                                                                                                                                                                                                               |
 | ------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| lineage       | [4-step LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) → 2-step trajectory distillation → distribution matching → a spectral match against the teacher's own images on top                            |
+| lineage       | [4-step LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) → 2-step trajectory distillation → distribution matching → a spectral match against the teacher's own images → an artefact critic and detail terms → three critics taking turns      |
 | this release  | the current run's latest probed checkpoint, chosen by the 12-bucket sweep and by my own look at the renders; the run continues from it one recipe change at a time                                                            |
-| what it gives | usable two-step renders at every trained resolution: fine detail within a few percent of what the 4-step adapter carries, and a prompt-adherence judge that calls it a loss against the 8-step teacher on 6 of 45 renders — the same count the 4-step adapter scores. What it does not give is the teacher's own picture: see [Known issues](#known-issues) and [Measured against the teacher](#measured-against-the-teacher) |
+| what it gives | usable two-step renders at every trained resolution: fine detail at or just above the teacher's — from 1 megapixel up, closer to the teacher than the 4-step adapter — with the prompt's objects, counts, attributes and relations in place (a blind rubric finds 1 point missing out of 240). A judge asked which render follows the prompt better still prefers the 8-step teacher on 11 of 45, against 6 for the 4-step adapter, mostly on how a stylised prompt says things should look. What it does not give is the teacher's own picture: see [Known issues](#known-issues) and [Measured against the teacher](#measured-against-the-teacher) |
 
 ### Known issues
 
-The usual costs of two steps, in this order of how often they show: fine structure comes out soft or a few pixels out of register — feathers, skin texture, hair strands, signage, the surface of a distant object — most at 1280×1280 and above; a faint doubled contour on faces and limbs. **Faces depend on how much of the frame they occupy**: a portrait-sized face holds up, while small or distant faces — a crowd, a figure in a wide scene — lose their features first and can come out misshapen, since at that size a whole face is only a few of the blocks the model works in. On busy action or crowd scenes the composition can also repeat itself — an extra hand or held object, a figure duplicated in a crowd — where the 8-step and 4-step renders commit to one. On some prompts the composition itself differs from the 8-step render at the same seed: two steps is a shorter path from the same starting noise, so the image can settle on a different framing, pose or arrangement rather than a degraded version of the teacher's. Treat the teacher's render as a reference for quality, not as the picture two steps will reproduce. At the largest sizes a fine grain remains on the most textured subjects and skin reads slightly smoother and less saturated than the teacher's. Every one of these is being worked on; none is hidden in the sweeps or the examples.
+The usual costs of two steps, in this order of how often they show. **Small subjects are the weak spot, people and objects alike**: a portrait-sized face or an object seen up close holds up, while small or distant subjects — faces in a crowd, a figure in a wide scene, the machines at the back of a room — can come out ghosted, smeared or misshapen, since at that size a whole subject is only a few of the blocks the model works in. Fine structure can come out soft or a few pixels out of register — feathers, hair strands, signage, the surface of a distant object — most at 1280×1280 and above, and a faint doubled contour can show on limbs. On stylised prompts, **how the prompt says the image should look is followed less faithfully than what should be in it**: crisp anime linework, energetic brush strokes, the fingerprints in clay or a matte-painting finish come out closer to a generic rendering than the teacher's. On busy action or crowd scenes the composition can repeat itself — an extra hand or held object, a figure duplicated in a crowd — where the 8-step and 4-step renders commit to one. On some prompts the composition itself differs from the 8-step render at the same seed: two steps is a shorter path from the same starting noise, so the image can settle on a different framing, pose or arrangement rather than a degraded version of the teacher's. Treat the teacher's render as a reference for quality, not as the picture two steps will reproduce. Skin reads slightly smoother and less saturated than the teacher's, and colour overall runs a little under the teacher's at the largest sizes; freckles tend to gather into clusters rather than separate dots. A fine grain remains on the most textured subjects at the largest sizes, lighter than in the previous checkpoint. Every one of these is being worked on; none is hidden in the sweeps or the examples.
 
 ## How I got here
 
@@ -134,6 +147,79 @@ The recipe adjustments so far, each made on the measurement of the one before:
 6. a spectral match of the student's image against the teacher's own, on the whole latent and on decoded pixel
    windows — the term that finally reached the grain at large resolutions, after a critic, per-resolution weights and
    a filtered push had each been tried against it and retired
+7. the decoded-window spectral term given a much lighter hand — capped at a quarter of its earlier strength, which kept
+   the detail and took some grain out of flat areas
+8. the fake-score adapter updated four times per student step instead of twice, so it keeps up with what the student
+   currently makes; faint straight-line artefacts that had begun to appear on flat illustrated areas went away with it
+9. **an artefact critic** — a small head reading the frozen base model's own mid-network features, trained to tell the
+   teacher's finished images from the student's, with its push restricted to structure finer than 32 pixels and held
+   well below the distribution term — aimed at what the distribution term leaves behind: melted small faces and dense
+   detail smeared into blotches
+10. four detail terms together: the anchor counting the fine-detail part of its error twice; a one-sided floor that
+    stops the finest detail dropping below the level real photographs carry; a second-call target made by the teacher
+    finishing the image from the student's own first-call output, so the target shares the student's layout; and a
+    smoothness limit on the fake-score adapter, so the distribution push keeps pointing at detail rather than away from it
+11. **three critics taking turns** — the artefact critic joined by one whose real examples are half real photographs and one
+    weighted toward faces, one of them pushing on each step while the others keep training in between, because the three did
+    not fit in memory side by side
+
+Two further ideas were tried and taken back out: confining the distribution term to the second call's noise range, and
+a detail pyramid compared pixel by pixel against the teacher, which on inspection rewarded fading any detail it could
+not place exactly where the teacher had it.
+
+## chk00017464 vs chk00013663
+
+`chk00013663` (published 12 Sep 2026) was the first public checkpoint: distribution matching with the spectral match, and nothing
+yet aimed at the faults the distribution term leaves behind. `chk00017464` (14 Sep 2026) is 3,801 training samples later, and every
+one of those samples went to those faults — the grain and grid pattern at large sizes, small faces, dense detail — through five recipe
+changes, each kept only after its own look at the renders:
+
+1. the decoded-window spectral term brought down to a quarter of its strength
+2. the fake-score adapter updated four times per student step instead of twice
+3. the artefact critic, reading the frozen base model's own features
+4. four detail terms: the anchor counting fine-detail error twice, the one-sided photo floor, the teacher's finish of the student's
+   first call as the second call's target, and a smoothness limit on the fake adapter
+5. three critics taking turns: the artefact critic, a photo critic and a face critic
+
+Two ideas were tried and taken back out along the way: the distribution term confined to the second call's noise range, and a detail
+pyramid compared pixel by pixel against the teacher.
+
+**Detail at large sizes — the headline.** Every one of the 12 sweep resolutions × 15 prompts measured against the 8-step teacher, as in
+[Measured against the teacher](#measured-against-the-teacher). Above 1 megapixel the excess fine energy two steps used to put into
+images is roughly halved, and those sizes are now closer to the teacher than the [4-step adapter](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) on
+fine texture and on both grid bands:
+
+| 1.00 = the teacher | fine texture | 16-px band | 8-px band |
+| --- | --- | --- | --- |
+| 1280×1280 | 1.20 → **1.09** | 1.16 → **1.08** | 1.25 → **1.13** |
+| 1440×1280 | 1.21 → **1.08** | 1.19 → **1.09** | 1.28 → **1.14** |
+| 1440×1440 | 1.34 → **1.17** | 1.25 → **1.08** | 1.42 → **1.21** |
+
+Fine-texture energy and both grid bands come closer to the teacher at 10 of the 12 resolutions, the blur-invariant ghosting index at
+10 of 12, the micro-ghost index at all 12, and the grain in flat areas — skies, walls, out-of-focus backgrounds — at 11 of 12
+(1440×1440: 1.84× the teacher's → 1.57×, median of the 15 prompts). It shows where it should: hair renders as more individual strands, and the token-grid
+texture that read as grain on birds, food and foliage at 1440×1440 is lighter.
+
+**Faces and skin.** Freckles on the test portrait gather into lighter, more dot-like clusters than before — better, not yet the
+teacher's separate dots — and eyes look about the same: clean irises, lashes softer than the teacher's.
+
+**What did not improve.** The judge that asks which of two renders follows the prompt better prefers the 8-step teacher on 11 of 45
+against 6 before, mostly on how a stylised prompt says the picture should look: crisp linework, energetic brush strokes, the texture
+of clay. The blind rubric that checks each render alone for the prompt's objects, counts, attributes and relations moves from 0 to 1
+point missing out of 240, and on the same 15 fresh prompts the previous checkpoint was measured on, the two checkpoints come out level
+(1 win, 13 ties, 1 loss). Colour runs slightly lower (saturation 0.93× the teacher's across the sweep, from 0.94×; 0.86× at
+1440×1440), and at 512×512 the two checkpoints are level. Both are what the next recipe changes target.
+
+| axis | `chk00013663` | `chk00017464` |
+| --- | --- | --- |
+| fine texture vs the teacher, 1280² / 1440² | 1.20 / 1.34 | **1.09 / 1.17** |
+| 16-px grid band, 1280² / 1440² | 1.16 / 1.25 | **1.08 / 1.08** |
+| grain in flat areas, sweep median | 1.36× | **1.26×** |
+| distance to the teacher, sweep mean | 0.413 | **0.406** |
+| judge prefers the teacher (of 45) | **6** | 11 |
+| blind adherence rubric, points missing of 240 | **0** | 1 |
+| saturation vs the teacher, sweep mean | **0.94×** | 0.93× |
+| training samples in the 2-step stages | 13,663 | 17,464 |
 
 ## Measured against the teacher
 
@@ -148,36 +234,41 @@ teacher):
 
 | bucket | fine texture | 16-px band | 8-px band | 4-step LoRA (same three) | stock 2-step, fine texture |
 | --- | --- | --- | --- | --- | --- |
-| 512×512 | 1.00 | 1.01 | 1.05 | 1.05 · 1.05 · 1.07 | 0.57 |
-| 768×1024 | 1.04 | 1.01 | 1.10 | 1.05 · 1.05 · 1.08 | 0.39 |
-| 1024×1024 | 1.11 | 1.12 | 1.15 | 1.11 · 1.17 · 1.19 | 0.39 |
-| 1280×1280 | 1.20 | 1.16 | 1.25 | 1.20 · 1.18 · 1.22 | 0.41 |
-| 1440×1440 | 1.34 | 1.25 | 1.42 | 1.24 · 1.23 · 1.32 | 0.42 |
+| 512×512 | 1.02 | 1.02 | 1.05 | 1.05 · 1.05 · 1.07 | 0.57 |
+| 768×1024 | 1.02 | 0.98 | 1.07 | 1.05 · 1.05 · 1.08 | 0.39 |
+| 1024×1024 | 1.06 | 1.07 | 1.12 | 1.11 · 1.17 · 1.16 | 0.39 |
+| 1280×1280 | 1.09 | 1.08 | 1.13 | 1.20 · 1.18 · 1.22 | 0.41 |
+| 1440×1440 | 1.17 | 1.08 | 1.21 | 1.24 · 1.23 · 1.32 | 0.42 |
 
 Two steps without the adapter carry **less than half** the teacher's fine detail at every size. With it, the detail
-lands within a few percent of what the 4-step adapter carries — at 1440×1440 slightly above it.
+sits at or just above the teacher's everywhere — and from 1 megapixel up it is closer to the teacher than the 4-step
+adapter, which carries more excess fine energy there.
 
 **Prompt adherence, judged.** A vision-language judge is shown the teacher's render and this LoRA's for the same prompt,
 in both orders, and asked which follows the prompt better; a loss means the teacher was preferred both times:
 
 | bucket | wins | ties | losses |
 | --- | --- | --- | --- |
-| 512×512 | 3 | 10 | 2 |
-| 1280×1280 | 1 | 12 | 2 |
-| 1440×1440 | 2 | 11 | 2 |
-
-Six losses out of 45 — the same count the [4-step LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) scores against the same teacher. On 15 prompts drawn
-fresh from the training prompt bank and never rendered before, the same judge returned 1 win, 13 ties, 1 loss.
-
-**Checked for the damage this kind of training can do.** Saturation sits at 1.01× the teacher's at 768×1024 and 0.91×
-at the larger sizes; edge detail 0.92–0.96×; skin texture inside detected faces 0.96× at 768×1024 and 0.85× at
-1440×1440, with skin saturation 0.93× and 0.81×. The honest reading of those last two: **skin is the softest and least
-saturated part of this adapter's output, and it gets softer as the render gets larger.** Fine detail in flat regions —
-skies, walls, out-of-focus backgrounds — runs 1.65–1.88× the teacher's, which is where two steps put grain that eight
-steps do not.
-
-**Distance to the teacher**, as a plain pixel measure, is 0.39–0.42 at every size against the [4-step LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA)'s
-0.32–0.34. That gap is what two model calls cost instead of four: the image is a good render of the prompt, but it is
+| 512×512 | 1 | 10 | 4 |
+| 1280×1280 | 1 | 10 | 4 |
+| 1440×1440 | 0 | 12 | 3 |
+
+Eleven losses out of 45, where the [4-step LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) scores six against the same teacher. They gather on
+stylised prompts and on how a prompt says the picture should look — crisp linework, energetic brush strokes, the texture of
+clay — more than on what should be in it: a blind rubric that scores each render on its own against the prompt's objects, counts,
+attributes and relations, with the teacher scored identically, finds 1 point missing out of 240. On 15 prompts drawn fresh from the
+training prompt bank and never rendered before, the judge returned 0 wins, 10 ties, 5 losses; on another 15 fresh prompts, 1 win,
+13 ties, 1 loss.
+
+**Checked for the damage this kind of training can do.** Saturation sits at 0.99× the teacher's at 768×1024 and 0.89×
+at the larger sizes; edge detail 0.93–0.96×; skin texture inside detected faces 1.07× at 768×1024 and 0.85× at
+1440×1440, with skin saturation 0.92× and 0.77–0.85× at the larger sizes. The honest reading of those last two: **skin
+is the softest and least saturated part of this adapter's output at large sizes, and colour overall runs a little under
+the teacher's there.** Fine detail in flat regions — skies, walls, out-of-focus backgrounds — runs 1.53–1.85× the
+teacher's, which is where two steps put grain that eight steps do not.
+
+**Distance to the teacher**, as a plain pixel measure, is 0.39–0.44 at every size against the [4-step LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA)'s
+0.30–0.37. That gap is what two model calls cost instead of four: the image is a good render of the prompt, but it is
 not the teacher's render of it — see [Known issues](#known-issues).
 
 ## Usage
@@ -190,7 +281,7 @@ not the teacher's render of it — see [Known issues](#known-issues).
 | guidance / CFG | **0.0** (Turbo is CFG-free; do not enable it)   |
 | timestep shift | **mu = 1.15**, fixed (Turbo's deployment shift) |
 
-The 2 sampling sigmas are Turbo's own deployment grid: `[1.0, 0.51284]` — the first and the middle
+The 2 sampling sigmas are Turbo's own deployment grid: `[1.0, 0.7595]` — the first and the middle
 of the 4-step grid, so the model is evaluated at two points it already knows.
 
 ## Inference with diffusers
@@ -309,18 +400,27 @@ the model already loaded, so the numbers are the render itself and not a model l
 
 | configuration | denoising | per model call | GPU peak |
 | --- | --- | --- | --- |
-| Krea 2 Turbo — 8 steps (the reference) | **79.9 s** | 9.99 s | 25.2 GiB |
-| Krea 2 Turbo — 2 steps, no LoRA | 20.2 s | 10.12 s | 25.2 GiB |
-| **Krea 2 Turbo — 2 steps + this LoRA** | **20.9 s** | 10.47 s | 25.2 GiB |
+| Krea 2 Turbo — 8 steps (the reference) | **81.4 s** | 10.2 s | 25.2 GiB |
+| Krea 2 Turbo — 2 steps, no LoRA | 20.4 s | 10.2 s | 25.2 GiB |
+| **Krea 2 Turbo — 2 steps + this LoRA** | **19.5 s** | 9.8 s | 25.2 GiB |
 
-**Denoising is 3.8× faster than the 8-step reference** — two model calls instead of eight. The adapter itself costs about
-**3.5% more per call** (10.47 s against 10.12 s) and no measurable extra memory, because a rank-64 low-rank product is
-small beside the transformer it is added to; that overhead is paid twice here instead of the four or eight times a longer
-schedule would pay it.
+**Denoising is 4.2× faster than the 8-step reference** — two model calls instead of eight. The adapter's own cost per call did
+not show up in this measurement: the runs with it came in marginally faster than those without, which is measurement noise, not a
+speed-up. A rank-64 low-rank product is small beside the transformer it is added to, and it adds no measurable memory.
 
 Denoising is the part the step count changes. What a complete render costs on top of it — encoding the prompt, decoding
 the latent, writing the file — is the same whether you run two steps or eight, and it depends on your pipeline, so the
-end-to-end figure on your machine will sit below 3.8× and rise toward it as the render gets larger.
+end-to-end figure on your machine will sit below 4.2× and rise toward it as the render gets larger.
+
+**By resolution.** The two model calls of this LoRA's own sweep renders on the same machine (median of the 15 test prompts per
+size; sweep renders run one at a time, not the controlled measurement above):
+
+| resolution | denoising (2 calls) | resolution | denoising (2 calls) |
+| --- | --- | --- | --- |
+| 512×512 | 6.4 s | 1024×1024 | 20.4 s |
+| 512×768 / 768×512 | 8.9 s / 9.0 s | 1280×960 / 960×1280 | 23.2 s / 24.0 s |
+| 768×768 | 11.9 s | 1280×1280 | 32.2 s |
+| 768×1024 / 1024×768 | 14.9 s / 15.1 s | 1440×1280 / 1440×1440 | 34.8 s / 39.8 s |
 
 ## LoRA strength
 
@@ -398,9 +498,9 @@ In every case the tensors are the same 456 bf16 matrices; only the names around
 **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
+The student makes two calls, at σ = 1.0 and σ = 0.7595 — 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
+from pure noise lands _exactly_ on the flow-matching interpolant at σ = 0.7595 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.
 
@@ -409,7 +509,8 @@ freshly noised copy: the frozen teacher, and a second small adapter on the same
 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.
+rank 32, starts as an exact copy of the teacher, updates four times per student step — often enough to keep up with a
+student that is still changing — 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
@@ -420,24 +521,24 @@ moved so far so fast.
 and under-fits fine structure there, and a per-pixel normaliser lands harder as pixel counts grow. A full 12-bucket
 sweep of the first distribution-matching checkpoint located the problem at 1 megapixel and above (fine-texture energy
 1.4–1.8× the teacher's at the five largest buckets); scaling the push per bucket from that measurement was tried and did
-not hold, and the spectral match replaced it. The same sweep of the current checkpoint, live weights, fixed seed, 15
-prompts per bucket, every image measured against the teacher's render of the same prompt and seed (in brackets: the
+not hold, and the spectral match replaced it. The same sweep of the published checkpoint, its running-average weights, fixed
+seed, 15 prompts per bucket, every image measured against the teacher's render of the same prompt and seed (in brackets: the
 first distribution-matching checkpoint on the same prompts):
 
 | bucket    | fine texture vs the teacher | 16-px grid band | 8-px grid band | distance to the teacher |
 | --------- | --------------------------- | --------------- | -------------- | ----------------------- |
-| 512x512   | 1.00 (1.16)           | 1.01 (1.23)     | 1.05 (1.23)    | 0.42 (0.44)             |
-| 512x768   | 1.01 (1.23)           | 1.08 (1.37)     | 1.06 (1.33)    | 0.39 (0.41)             |
-| 768x512   | 1.04 (1.22)           | 1.08 (1.37)     | 1.09 (1.26)    | 0.45 (0.47)             |
-| 768x768   | 1.09 (1.43)           | 1.06 (1.46)     | 1.15 (1.50)    | 0.40 (0.43)             |
-| 768x1024  | 1.04 (1.36)           | 1.01 (1.38)     | 1.10 (1.42)    | 0.39 (0.43)             |
-| 1024x768  | 1.04 (1.35)           | 1.03 (1.40)     | 1.07 (1.38)    | 0.43 (0.46)             |
-| 1024x1024 | 1.11 (1.47)           | 1.12 (1.55)     | 1.18 (1.56)    | 0.42 (0.44)             |
-| 1280x960  | 1.09 (1.50)           | 1.07 (1.59)     | 1.16 (1.56)    | 0.42 (0.42)             |
-| 960x1280  | 1.20 (1.60)           | 1.16 (1.63)     | 1.30 (1.70)    | 0.41 (0.43)             |
-| 1280x1280 | 1.20 (1.65)           | 1.16 (1.68)     | 1.25 (1.69)    | 0.42 (0.44)             |
-| 1440x1280 | 1.20 (1.59)           | 1.19 (1.66)     | 1.28 (1.66)    | 0.40 (0.42)             |
-| 1440x1440 | 1.34 (1.81)           | 1.25 (1.79)     | 1.42 (1.89)    | 0.40 (0.41)             |
+| 512x512   | 1.02 (1.16)           | 1.02 (1.23)     | 1.05 (1.23)    | 0.41 (0.44)             |
+| 512x768   | 1.01 (1.23)           | 1.06 (1.37)     | 1.06 (1.33)    | 0.39 (0.41)             |
+| 768x512   | 1.02 (1.22)           | 1.08 (1.37)     | 1.04 (1.26)    | 0.44 (0.47)             |
+| 768x768   | 1.09 (1.43)           | 1.05 (1.46)     | 1.14 (1.50)    | 0.40 (0.43)             |
+| 768x1024  | 1.02 (1.36)           | 0.98 (1.38)     | 1.07 (1.42)    | 0.39 (0.43)             |
+| 1024x768  | 1.01 (1.35)           | 0.98 (1.40)     | 1.04 (1.38)    | 0.43 (0.46)             |
+| 1024x1024 | 1.06 (1.47)           | 1.07 (1.55)     | 1.12 (1.56)    | 0.41 (0.44)             |
+| 1280x960  | 1.03 (1.50)           | 1.04 (1.59)     | 1.07 (1.56)    | 0.40 (0.42)             |
+| 960x1280  | 1.10 (1.60)           | 1.05 (1.63)     | 1.19 (1.70)    | 0.40 (0.43)             |
+| 1280x1280 | 1.09 (1.65)           | 1.08 (1.68)     | 1.13 (1.69)    | 0.41 (0.44)             |
+| 1440x1280 | 1.08 (1.59)           | 1.08 (1.66)     | 1.14 (1.66)    | 0.40 (0.42)             |
+| 1440x1440 | 1.17 (1.81)           | 1.08 (1.79)     | 1.21 (1.89)    | 0.39 (0.41)             |
 
 ### The spectral match
 
@@ -455,8 +556,60 @@ latent every sample, and on one random 256-pixel window decoded through the VAE
 judged too. The comparison is two-sided, so too much fine energy and too little are both penalised; a blurred image
 does not satisfy it. Its gradient is added to the distribution push and capped per sample as a fraction of it, so it
 refines rather than takes over. At its first strength it brought every resolution closer to the teacher's spectrum
-without touching adherence, layout or variety; raised, it began closing the grain on the hardest subjects too, and that is
-the setting the current run trains with.
+without touching adherence, layout or variety; raised, it began closing the grain on the hardest subjects too. The whole-latent comparison trains at that
+strength; the decoded window was later brought down to a quarter of it, which kept the detail and removed some grain.
+
+### The artefact critic
+
+Distribution matching improves what the fake adapter can see, and the fake adapter learns from the student's own
+images — so where the student smears something, the fake learns the smear and the push stops correcting it. The two
+places that shows most are small faces, which come out melted, and dense content such as the goods on a market stall or
+the shelves of a shop seen through its window, which comes out as coloured blotches.
+
+A critic breaks that loop by looking at finished images instead. It is a small head on the frozen base model's
+mid-network features — every adapter switched off, the forward pass stopped halfway, so no adapter can learn to fool
+it and nothing it learns leaks into the fake adapter. Its real examples are the teacher's own finished images for
+other prompts at the same resolution; its fake examples are the student's final images. Both are lightly re-noised
+first, in the low-noise range where fine structure lives, and the critic reads them with an empty prompt, so it judges
+only whether the structure looks like the teacher's. Its push on the student is filtered to periods finer than 32
+pixels — below that it can rebuild a face or a shelf, above it it could move layout, colour or pose, which it must not
+— and capped at a quarter of the distribution term's strength per sample. Lazy gradient regularisation and spectral
+normalisation keep the head from overshooting. On the largest steps, where memory is tightest, the critic and the
+teacher's finishing pass below take turns instead of sharing a step.
+
+### Critics in turn
+
+One critic holds one idea of what is wrong. The artefact critic was joined by two more on the same frozen mid-network
+features and under the same rules — lightly re-noised inputs, an empty prompt, a push that is filtered and capped — each
+aimed at a different fault:
+
+- **A photo critic.** Half of its real examples are real photographs and half the teacher's finished images, so it learns
+  what fine texture looks like in a photograph as well as in the teacher's rendering of one. Its push is filtered to
+  periods finer than 24 pixels and held lower than the artefact critic's, because photographs carry grain the teacher
+  does not.
+- **A face critic.** Its real examples are the teacher's finished images of prompts with faces, the face regions counted at
+  full weight and the rest at half, so its push concentrates on what small and mid-sized faces lose first.
+
+All three on every step do not fit in 24 GB, so they take turns: on each step one critic pushes, and on alternate steps the
+others train so none goes stale before its turn comes back. The two new heads started from the artefact critic's weights and
+trained on their own before they were allowed to push.
+
+### Detail terms
+
+Four smaller terms sit on top, each capped relative to the distribution term so none of them can take over:
+
+- **A detail-weighted anchor.** The trajectory regression counts the fine-detail part of its error — everything finer
+  than 32 pixels — twice, so the anchor stops tolerating softness it used to average away.
+- **A one-sided photo floor.** On the decoded window, the student's energy at periods of 3–10 pixels may not fall below
+  the teacher's plus the margin real photographs carry over it at those scales. That margin is measured once from a
+  pool of real photographs and clamped, and the term only ever pushes upward to that floor, never past it — so it
+  lifts detail that is missing without adding grain that is not.
+- **The teacher's finish as a target.** Every second step, the teacher itself runs its remaining steps starting from
+  the student's own first-call output. The result is a finished image that shares the student's layout, and the second
+  call is pulled gently toward it — a target that lines up with what the student actually drew, where the recorded
+  trajectory might have drawn something else.
+- **A smoothness limit on the fake adapter.** The fake adapter's fine-detail energy is kept below the teacher's at the
+  point where the distribution term is measured, so the difference between the two keeps pointing toward detail.
 
 ## What the LoRA touches
 
@@ -469,8 +622,10 @@ of all 28 transformer blocks, plus the four global linears — `time_embed.linea
 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 spectral
-match reads 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
-not used in the current run.
+match and the artefact critic read 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) enter in two places: as one precomputed statistic — how much fine-detail energy they carry at 3–10
+pixels relative to the teacher, clamped — which sets the photo floor, and as half of the photo critic's real examples. Only that
+critic's head sees them; the student and the fake adapter never do, and receive only its filtered, capped push.
 
 ## Resolutions
 
@@ -541,22 +696,23 @@ The resolutions are exactly the training buckets listed under [Resolutions](#res
 ## 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 spectral term 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 21.4 GB at 1440×1440, of 24. The price of
-the objective is throughput: **188 training samples an hour** measured over a complete 10-hour run, against the
-[4-step recipe](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA)'s 470 — two and a half times the
-cost per sample, and so far a small fraction of the samples.
+host memory above 0.3 megapixels; the student, the fake adapter, the spectral and detail terms, the critic and the teacher's finishing pass 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 and every critic pushing reserves about 22.4 GB at 1440×1440,
+of 24. The price of the objective is throughput: **about 107 training samples an hour** measured over the recipe's complete
+run, against the [4-step recipe](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA)'s 470 — more than four times the cost per sample, and so far a small
+fraction of the samples.
 
 **Where that cost comes from.** Distribution matching is simply a heavier objective than trajectory distillation.
 The 4-step project's recipe compared the student's own output with a teacher state that had already been recorded to
 disk, so a training step was one student pass plus a small adversarial head. Here every step also needs the *score*
 of two models at a freshly noised point: the frozen teacher's, and a second adapter's that is being trained
-alongside to imitate the student — and that second adapter takes two optimiser steps of its own per student step.
-A third term then decodes part of the image out of the latent to compare its texture with the teacher's, which costs
-another pass through the decoder.
+alongside to imitate the student — and that second adapter takes four optimiser steps of its own per student step.
+The spectral and detail terms decode part of the image out of the latent to compare its texture with the teacher's,
+each critic reads half the network twice more, and every second step the teacher finishes the image from the student's
+first call.
 
-Counted in whole model runs per training sample, the difference is roughly **two there against seven here**. None of
+Counted in whole model runs per training sample, the difference is roughly **two there against about a dozen here**. None of
 that difference is the teacher generating anything: its renders were recorded once for the 4-step project and are
 read from disk by both. The extra work is the objective itself, and it bought the only thing that mattered. Run at
 two steps, the 4-step project's recipe reached a point where more training changed nothing: the measurements sat
@@ -570,7 +726,8 @@ At regular intervals, both the live weights and their running average are pulled
 15 fixed prompts across four resolutions (512×512, 768×1024, 1280×1280, 1440×1440); milestone checkpoints get the same
 render at all 12 buckets, which is where the per-resolution table above comes from. Every image is measured against the
 teacher's render of the same prompt and seed: distance, fine-texture energy, the 16-pixel and 8-pixel grid bands, skin
-and flat-region grain, saturation, faces cut out at 1:1, a graded judge, a pairwise preference against the teacher, and a blind rubric that
+and flat-region grain, saturation, faces cut out at 1:1, fixed content windows (small faces in a crowd, shop interiors
+seen through their windows), straight-line artefacts, a graded judge, a pairwise preference against the teacher, and a blind rubric that
 scores each render on its own against the prompt's objects, counts, attributes and relations, with the teacher scored
 identically. Fifteen prompts drawn fresh from the prompt bank, never rendered before, are judged the same way at every
 checkpoint. Latent distances — the held-out chord gap and the two-step rollout
@@ -928,12 +1085,19 @@ _Still out-of-spec and still edge case preview-only — this is one step, not th
 
 ## What's next
 
-Training continues from the released checkpoint, one recipe change at a time, each kept only if the pictures do not
-degrade at any resolution — aiming at the best quality two steps can give, not at matching the [4-step LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA). The
-levers on the table are the ones the remaining defects point at: fine structure that comes out soft or a few pixels
-out of register at the largest sizes, small faces in crowds, signage. A better checkpoint replaces this one when the
-sweeps and I visually agree, the same discipline as the [4-step adapter](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA); until then the
-[4-step adapter](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) remains the recommendation for quality renders, and this one is the fast preview.
+Training continues from this checkpoint, one recipe change at a time, each kept only if the pictures do not degrade at any
+resolution — aiming at the best quality two steps can give, not at matching the [4-step LoRA](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA). The next step
+is already running, and it goes after what this checkpoint still gets wrong: a critic that judges the first call against the
+teacher's own intermediate state from the same noise, so a first call that blends two layouts is caught where the blend
+happens; a critic weighted toward wherever the teacher put fine detail; the photo critic and the photo floor limited to
+photographic prompts, so illustration, anime and 3D renders are no longer pulled toward photographic grain; detail held to the
+teacher region by region, with a ceiling as well as a floor; a focus on the eyes, nose and lips of faces so they sharpen while
+skin stays the teacher's; a colour floor, so colour at large sizes stops falling below the teacher's; and a more even mix of
+resolutions. After it comes prompt adherence — a critic that learns whether an image belongs to its own prompt, and stylised
+prompts drawn more often — held to the rule that none of it may cost the sharpness this checkpoint gained. A better checkpoint
+replaces this one when the sweeps and I visually agree, the same discipline as the
+[4-step adapter](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA); until then the [4-step adapter](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-4step-LoRA) remains the recommendation for quality renders, and this
+one is the fast preview.
 
 ## License
 

krea2_turbo_2step_rank_64_lora.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:59306ece2d3d55186c1498752a04816f19d9acb195d219e330e1f4c59228b614
-size 438160888
+oid sha256:cc7a7d65ca04070b2aa1dbf60b3cd0cf210fde2a93d4d7d4fc18bdadf23060d3
+size 438161144

krea2_turbo_2step_rank_64_lora_checkpoint_info.md

@@ -1,13 +1,13 @@
 # Which checkpoint is this?
 
 `krea2_turbo_2step_rank_64_lora.safetensors` and `krea2_turbo_2step_rank_64_lora_comfyui.safetensors` in this folder are
-**chk00013663** — the same weights as `_archive/checkpoints/krea2_turbo_2step_rank_64_lora_chk00013663.safetensors` (and its
+**chk00017464** — the same weights as `_archive/checkpoints/krea2_turbo_2step_rank_64_lora_chk00017464.safetensors` (and its
 `_comfyui` twin). The pair here is updated in place whenever a better checkpoint ships; the archive
 keeps every one that did. The same checkpoint id is in each file's safetensors metadata (`checkpoint`).
 
 | file | SHA-256 | size |
 | --- | --- | --- |
-| `krea2_turbo_2step_rank_64_lora.safetensors` | `59306ece2d3d55186c1498752a04816f19d9acb195d219e330e1f4c59228b614` | 418M |
-| `krea2_turbo_2step_rank_64_lora_comfyui.safetensors` | `0887ac53a78a015c22c6fc6dfa2f76e9fca7e8c19ac074d65ad12146cbacb999` | 418M |
+| `krea2_turbo_2step_rank_64_lora.safetensors` | `cc7a7d65ca04070b2aa1dbf60b3cd0cf210fde2a93d4d7d4fc18bdadf23060d3` | 418M |
+| `krea2_turbo_2step_rank_64_lora_comfyui.safetensors` | `5c02cac5de27dcb40ea98caee5fd5668ac1a928a105a1fb29c02eb9b0b1e5862` | 418M |
 
-Updated: 12 Sep 2026
+Updated: 14 Sep 2026

krea2_turbo_2step_rank_64_lora_comfyui.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0887ac53a78a015c22c6fc6dfa2f76e9fca7e8c19ac074d65ad12146cbacb999
-size 438141424
+oid sha256:5c02cac5de27dcb40ea98caee5fd5668ac1a928a105a1fb29c02eb9b0b1e5862
+size 438141680