card: step_25000 samples + honest status

README.md

@@ -16,7 +16,7 @@ tags:
 
 _MOSS-Image pruned to fit a 16 GB card: a 30-block DiT instead of 42, a 0.8B text encoder with a 158M adapter instead of the 15.6 GB 9B one._ Text-to-image at 1024 px, **12.4 GB peak**, one consumer GPU.
 
-![five portraits, 768x1152](samples/hero.jpg)
+![four samples: cafe portrait, armoured figure, fashion shot, mouse knight](samples/hero.jpg)
 
 **This is an alpha.** It was assembled in two days on a single **RTX 4080 (16 GB)** and it is **undertrained**: see [Status](#status). Treat it as a working proof that the base model survives a 26.7% cut and a 16× smaller encoder, not as a finished model.
 
@@ -28,7 +28,7 @@ _MOSS-Image pruned to fit a 16 GB card: a 30-block DiT instead of 42, a 0.8B tex
 
 **scheduler** — `MOSSImageScheduler`, flow-matching Euler, static shift 5.0.
 
-**resolution** — 1024 by default, any positive multiple of 16; 3:4 and 4:3 work (the samples below are 768x1152).
+**resolution** — 1024 by default, any positive multiple of 16; 3:4 and 4:3 work (the samples below are 768x1152, 768x1024, 864x1152 and 1152x1152).
 
 **peak VRAM** — **12.4 GB** at 1024 px / 30 steps, one forward pair per step, no offload. 9.7 GB at 768x1152 when the encoder is freed before the DiT runs; ~1.15 it/s on a 4080.
 
@@ -48,15 +48,16 @@ A mossling is a small moss. The base model is MOSS-Image, this is the same plant
 
 ## Examples
 
-Every image below comes from this folder: 768x1152, 50 steps, CFG 3, seed 7. The adapter computes the condition first and is freed before the DiT runs.
+Every image below comes from this folder: 50 steps, CFG 3, seed 7, each at its own aspect ratio. The adapter computes the condition first and is freed before the DiT runs. Only the first prompt is mine; the other three are copied verbatim from rows of the fine-tune set, resolutions included.
 
 | | prompt |
 | --- | --- |
 | ![cafe](samples/cafe_window_768x1152.png) | A young woman with curly red hair sits by the window of a sunlit cafe, holding a ceramic cup with both hands, warm morning light on her face, soft shadows, shallow depth of field, 85mm portrait photograph, natural skin texture, detailed eyes |
-| ![cliff](samples/red_dress_cliff_768x1152.png) | A woman in a flowing red dress stands on a windswept coastal cliff at golden hour, the fabric moving in the wind, dramatic clouds and wild grass, full body, wide shot, cinematic photography, warm rim light, 50mm |
-| ![fisherman](samples/fisherman_dawn_768x1152.png) | An elderly fisherman with a weathered face and a white beard mends a fishing net on a wooden boat at dawn, misty harbor with moored boats behind him, muted colors, documentary photograph, 50mm, fine detail in the hands and rope |
-| ![neon](samples/neon_rain_768x1152.png) | A smiling couple walks hand in hand along a rainy city street at night, colorful neon signs reflected in the wet asphalt, umbrellas and passing cars, cinematic photograph, 35mm, shallow depth of field, bokeh |
-| ![ballet](samples/ballet_studio_768x1152.png) | A ballet dancer in a white tutu stretches at the barre of an empty studio, tall windows with warm afternoon light, wooden floor, dust in the sunbeams, full body, film photograph, soft grain, 35mm |
+| ![wolverine](samples/wolverine_armor_1152x1152.png) | This highly detailed CGI rendering captures the iconic character Wolverine from Marvel Comics in his futuristic armor suit. The costume features intricate blue and gold plates, with sharp claws protruding from both hands, emphasizing his signature weapon. The background is dimly lit, highlighting the armored figure's muscular physique and determined expression under a glowing helmet. |
+| ![bazaar](samples/bazaar_leather_768x1024.png) | The image shows a woman seated in a vibrant red chair, with her legs crossed. She is wearing a stylish black leather pajama set with white trim, and a pair of black goggles rests on her head. She is also wearing red lipstick and a watch on her left wrist. In her lap, she cradles a small white dog wearing a black cap. The backdrop is a stark white, with the word "BAZAAR" written in bold black letters at the top. LBD Gigi and Bella Hadid |
+| ![mouse](samples/mouse_knight_864x1152.png) | A cartoon of a grey-furred female mouse with golden eyes, her body retaining all the natural, unmistakable features of a mouse—small, round ears, a long, twitching tail, and paws with delicate fingers. She wears simple beige clothing suited for travel, with small, practical bronze armor pieces tailored perfectly for her size. A long, curved sword rests on her shoulder, the hilt grasped firmly in her paw as her stance remains low to the ground, reflecting her nimbleness. The background is an arid desert, with the ruins of an ancient temple in the distance, weathered by time and partially overtaken by shifting sands. |
+
+The mouse row's caption continues with twelve random words ("Florgnax qwermbly …"); they are dropped here and were kept out of the prompt.
 
 ## Usage
 
@@ -89,11 +90,11 @@ Training and compression scripts are in `src/` (`src/README.md`): `window.py` +
 
 Honest list, this is what "alpha" means here:
 
-* **The fine-tune is short.** The DiT in this folder is the fused 30-block build after ~10 000 optimizer steps on ~21 000 captioned images; the run was then stopped — no convergence, no lr schedule, no EMA. Expect visible roughness on hands, complex geometry and crowded scenes.
+* **The fine-tune is short and flat.** The DiT in this folder is the fused 30-block build after one epoch's worth of fine-tuning: 39 025 captioned rows, stopped at **25 000 of 39 025 steps** (lr 1e-5 cosine to 1e-6, 200-step warmup, no EMA). It did not converge. Re-shot on five fixed prompts with a fixed seed, 25 000 steps are not visibly better than 2 000 and add small-structure regressions on top: an extra arm in the cafe portrait, fused limbs in a ballet shot, mushier faces in a crowded neon street. Stopping was a decision, not an accident — more steps were trading crispness for smoothness. Expect visible roughness on hands, complex geometry and crowded scenes.
 * **Not a full re-train.** The base model's own training is inherited; only the block merges and the short fine-tune above are new.
-* **Text inside the image degrades** — signs and JSON layouts come out garbled. Two of the fused blocks carried that ability, and one block plus a low-rank patch does not reproduce it.
+* **Text inside the image degrades** — signs and JSON layouts come out garbled, though one large word can land (the `BAZAAR` sample). Two of the fused blocks carried that ability, and one block plus a low-rank patch does not reproduce it.
 * **JSON layout prompts are the weakest case** for the adapter as well (cosine 0.756): it was trained on captions, not on layout instructions — the two limitations stack on sign/layout prompts.
-* **One seed per prompt in the samples.** Quality varies with seed; the images above are not cherry-picked from hundreds, they are the first five prompts tried.
+* **One seed per prompt in the samples.** Quality varies with seed, and the four images above are one draw each, seed 7, nothing re-rolled: three prompts are random rows of the fine-tune set (resolutions included), the cafe one is mine. They are picked to show what the model does, not the best of hundreds.
 * **English prompts**, as in the dataset behind the adapter.
 
 Quality against the base model, measured on the pruned-but-not-fine-tuned build:
@@ -116,7 +117,7 @@ text_encoder/            Qwen3.5-0.8B, frozen
 adapter/                 adapter v1: adapter.safetensors (158M head) + its own src/ and samples/
 vae/                     untouched AutoencoderKLFlux2, fp32
 scheduler/, tokenizer/   untouched
-samples/                 the five images above + their contact sheet
+samples/                 the four images above + hero.jpg
 src/                     compression and training scripts, plus the method write-up
 ```