BrutAIsm · Research table
Autophagy
What is left of Luchian when the model learns only from itself
An image model learned to imitate images by Ștefan Luchian, from 17 flower paintings in the MNAR collection. Then we let it feed on itself, in a loop: it generates the 17 images, a new model learns only from them, then generates again. Each generation feeds on the one before; no real painting comes in again. We expected the images to fade, to wear away like a photocopy of a photocopy. They did not fade. They hardened: stronger colours, thicker outlines, more contrast, then a gilt frame nobody asked for.
How to read. Each column is a generation. Each row is a different loop: how much the model relearns from its own images at every cycle, from a little (gentle) to a lot (harsh). All images start from the same point (the same seed) and are never used for training, so the changes come from the model.
Poppies and wheat ears in a jug · same starting point
| the painting | GENERATION 0 · trained on the 17 paintings | after 5 cycles of feeding on itself | |
|---|---|---|---|
 |  |  |
What a model learns
A model does not keep the paintings. It keeps what repeats in them: the red of the poppies, the jug, the brushwork, the signature. When we teach it only from its own images, it learns what repeats in those, and at the next generation, what repeats in those.
- 17 paintings
- model
- 17 generated images
- new model
- 17 generated images
- new model
- …
The models on this table
- Standard model The public model, untouched. It has never seen Luchian. The control.
- Generation 0 The starting point: the same model, with a thin layer (a LoRA) learned from the 17 real paintings. Every loop starts here.
I · Four loops
It does not fade. It hardens.
Four loops of five generations each, differing only in how much the model relearns from its own images at every cycle. In all four, colours saturate, contrast rises, outlines thicken, generation after generation. The gentle loop keeps the painting’s colours for a long time; the harsh one ends with cyan leaves and a purple ground. From the third or fourth generation a gilt frame appears at the edges, which nobody asked for. The “Luchian” signature survives almost to the end.
After 5 generations, the model fed on Ștefan Luchian’s flowers looks less and less like the originals. The model doesn’t go deeper. It collapses.
Poppies and wheat ears in a jug
The loop that also learns from cropped details moves in tighter on the flowers and softens toward pink: it is the only one in which outlines weaken instead of hardening, a different kind of decay.
| the painting | STANDARD MODEL · same prompt | GENERATION 0 | |||
|---|---|---|---|---|---|
| start | |  | |
| generation 1 | generation 2 | generation 3 | generation 4 | generation 5 | |
|---|---|---|---|---|---|
| gentlerelearns a little each cycle |  |  |  |  |  |
| mediumrelearns more |  |  |  |  |  |
| harshrelearns the most |  |  |  |  | |
| medium, with close-upsalso learns from crops of its own images |  |  |  |  |  |
Why do the colours harden?
A likely explanation, not tested here. When it generates, the model is pushed slightly away from an “ordinary” picture, the one with no words (see also the sheet “Appearance imitation”). This push is known to strengthen colour and contrast. The next model learns the strengthened images as if they were the paintings, and strengthens them again when it generates. A little each cycle, like a game of telephone, except that here the message is not lost: it thickens.
Poppies in a white vase
The painting’s warm ground turns green, then emerald. The white vase has ornaments from the first model on; the generations thicken them.
| the painting | STANDARD MODEL · same prompt | GENERATION 0 | |||
|---|---|---|---|---|---|
| start |  |  |  |
| generation 1 | generation 2 | generation 3 | generation 4 | generation 5 | |
|---|---|---|---|---|---|
| gentlerelearns a little each cycle |  |  |  |  |  |
| mediumrelearns more |  |  |  |  |  |
| harshrelearns the most |  |  |  |  |  |
| medium, with close-upsalso learns from crops of its own images |  |  |  |  |  |
White roses
The roses multiply until they fill the whole frame. In the harsh loop, the green turns fluorescent.
| the painting | STANDARD MODEL · same prompt | GENERATION 0 | |||
|---|---|---|---|---|---|
| start |  |  |  |
| generation 1 | generation 2 | generation 3 | generation 4 | generation 5 | |
|---|---|---|---|---|---|
| gentlerelearns a little each cycle |  |  |  |  |  |
| harshrelearns the most |  |  |  |  |  |
Carnations
| the painting | STANDARD MODEL · same prompt | GENERATION 0 | |||
|---|---|---|---|---|---|
| start |  |  |  |
| generation 1 | generation 2 | generation 3 | generation 4 | generation 5 | |
|---|---|---|---|---|---|
| gentlerelearns a little each cycle |  |  |  |  |  |
| harshrelearns the most |  |  |  |  |  |
II · Another loop
Each painting, another road
We repeated the loop on five other paintings, with a harsher recipe. The same recipe for all of them, but the paintings take different roads: some harden in their outlines, others in their texture, one blurs. The right-hand column is the control: without training, the word “luchian” means nothing to the model, so it grabs onto the inventory number and generates stone medallions, winter landscapes, engravings.
One recipe, five ways of getting lost.
| GENERATION 0 | generation 1 | generation 2 | generation 3 | generation 4 | STANDARD MODEL · “luchian, <number>” | |
|---|---|---|---|---|---|---|
| inv. 10.578Garoafe Outlines thicken; the white flowers are gone after the first generation. |  |  |  |  |  |  |
| inv. 7704Albăstrele The brushwork turns into glossy relief; the ground crumples like tin foil. |  |  |  |  |  |  |
| inv. 797Anemone The one that blurs: the flowers become pale discs, the vase green patches. |  |  |  |  |  |  |
| inv. 9766Maci în vas alb The yellow flowers turn red after the first generation, and the vase breaks into white fragments. |  |  |  |  |  |  |
| inv. 9292Flori părăluțe The vase changes its ornament from generation to generation; the flowers darken. |  |  |  |  |  |  |
prompt: “luchian” + the painting’s inventory number
A world that narrows
Luchian painted many of his flowers in the last years of his life, when illness kept him in his room, in an armchair or in bed. Friends brought him the flowers, and he set them in a vase and painted them. His world had narrowed to a room, but the flowers came from outside, different each time. Nobody brings the model flowers: all it receives are its own 17 images.
Conclusion
A model that feeds only on its own images does not forget slowly, toward grey. It exaggerates. What it kept from Luchian (the red of the poppies, the striped jug, the signature) repeats more and more emphatically, the images grow more alike, and the painting itself recedes. Researchers call it autophagy: models that feed only on their own data “go mad”.
Our loop is the extreme case: no real painting comes in again. On the internet, more and more images are generated and the next models learn from them too, but real images stay alongside the generated ones. Researchers disagree on how much that matters. Five cycles do not say where the road ends. They only say which way it goes.
The copy of a copy does not fade. It collapses into itself.
Methods
- SDXL 1.0 base
- LoRA trained on 17 paintings by Ștefan Luchian (MNAR)
- second loop: full fine-tune per cycle
- four loops: LoRA per cycle merged into the model
- similarity: CLIP
- references: Alemohammad et al., “Self-Consuming Generative Models Go MAD”, ICLR 2024
- Gerstgrasser et al., “Is Model Collapse Inevitable?”, 2024
- on guidance and oversaturation: Sadat et al., ICLR 2025
- September–October 2026
- full reports: BrutAIsm, research table