A Conditioning-Emergence Probe for Latent Diffusion Transformers
BlackMind's first image model: a latent diffusion transformer trained from scratch on CC12M whose prompt-conditioning is measured climbing from chance to 24% of the caption ceiling — plus a Muon optimiser win and a purpose-built CLIP metric that tracks prompt-conditioning the training loss does not capture.
In brief
BlackMind's first text-to-image model: a latent diffusion transformer trained from scratch on CC12M on two consumer GPUs. It learns to draw recognisable scenes and, measured against a controlled ceiling, learns to make its picture follow its caption — climbing from chance to 24% of ceiling — and a purpose-built CLIP metric tracks that conditioning directly, establishing which training interventions change the picture and which only move the loss. Every interim claim is held to its receipt under a pre-registered audit.
Key results
- A DiT-S (43.67M) trained from random initialisation memorises 64 fixed CC12M images and, sampling from pure noise, routes every caption to its own image 64 of 64 times (100% against a 1.6% chance baseline) in 7.52 minutes — validating the ported flow-matching core under the Muon+WSD harness with no downloaded weights, no LoRA, and no pretrained image prior.
- Trained for real, two Stage-1 seeds reach a held-out velocity-MSE of 0.1754 over ~15,800 steps (well below every short screen), and the model draws recognisable cars, kitchens, food and garments from held-out prompts it never saw in training.
- The headline positive result is measured conditioning emergence: on a controlled CLIP prompt-alignment probe the model climbs from ignoring its caption at step 2k (p=0.78), through a ~14,000-step plateau at ~16% of the ground-truth caption ceiling, to a breakout at 24% of ceiling by step 31.5k (paired gap +0.0483, CI95 [+0.0347,+0.0630], p<=0.001) — a from-scratch model measurably learning to make its picture depend on its prompt.
- The methodological contribution is a purpose-built perceptual metric: because held-out velocity-MSE — the only metric that ever gated an image decision here — is insensitive to text content (it reads ~0.2025 with real, shuffled, and absent captions, moving ~0.002 across an 8x parameter sweep), a paired CLIP prompt-alignment measure was built to track conditioning directly.
- A CLIP arm-similarity study then establishes which interventions change the picture: three of six pairs — including the entire learning-rate sweep — change it no more than a different seed does; the Muon-vs-AdamW winner (Muon leads loss 0.18333 vs 0.18609) is a perceptual near-tie; and quality-injection, the one intervention that demonstrably moves the image distribution, scores lower on the primary metric (+0.00347).
- A full classifier-free-guidance sweep sets the sampler to w=3.0 (the w=1 and w=5 alignment CIs are disjoint and w=3 is the optimum); guidance is primarily an aesthetic control (+70% saturation with flat sharpness); and every interim claim is held to its on-disk receipt under a pre-registered audit.
The Model and the Rig
This is M1, BlackMind's first image generator: a latent diffusion transformer trained from random initialisation on CC12M images, on two consumer graphics cards. There are no downloaded generator weights, no LoRA, and no pretrained image prior at any point in the run; the only frozen components are a VAE and a text encoder, used strictly as fixed instruments, and the transformer that does the drawing is ours, from scratch. The latents come from a frozen Wan2.1 VAE and the text pathway from a frozen Qwen3-1.7B; the objective is flow matching on a velocity-MSE loss in the rectified-flow formulation, with an SD3-style train-time timestep shift.
M1 is the first real training rung of a staged video-model program, and the image model comes first by design, because the cheapest way to prove a diffusion-transformer training pipeline works is to make it draw a still frame before asking it to make a moving one. The program runs on a falsifiable ladder: every milestone is a pre-registered gate, passed only by an audited receipt, and failing a gate is a result, not a delay. M1's gate is deliberately modest — our ~150-400M latent DiT, trained on our corpus, beats a frozen SD1.5-class open checkpoint on our suites under our judge protocol, at matched inference compute. It does not claim parity with modern frontier image models; it claims the pipeline works.
Two things make this worth writing up, and the first is that the model learned. The animation below steps through training, the same 16 held-out prompts sampled from pure noise at each checkpoint from step 1,984 to 31,520, and you can watch recognisable objects and scenes sharpen out of noise — cars, kitchens, food, garments, flowers — with distortion and garbled text, but unmistakably on the right track. The dependence of those pictures on their prompt is not asserted from eyeballing a grid: it is measured, and it climbs from chance to 24% of the ground-truth ceiling over the same training. The second, just as valuable, is the measurement discipline that framed it — sharp enough to build a perceptual metric for exactly the thing the training loss does not capture, which is what lets us state the 24%-of-ceiling result and stand behind it.

The Trainer, Validated
Before any ranking claim, the trainer earns the right to be trusted by memorising a tiny fixed set and routing text to it correctly. Sixty-four CC12M images were centre-cropped to 256x256, encoded once through the frozen Wan2.1 VAE, and overfit with zero text-dropout; over 1,000 steps the smoothed loss fell from 1.2444 to 0.00353 — a 352x drop, monotone, no spikes, no NaN, in 7.52 minutes of wall-clock.
The go/no-go was the sampling test. Drawing 64 images from pure noise, each conditioned on one of the 64 cached caption embeddings, every single sample is nearest in latent space to its own caption's ground-truth image: 64 of 64, 100%, against a 1.6% chance baseline. We ran the first-skeptic check — this is not a decode-the-ground-truth bug, because the samples start from pure noise and the mean latent distance to the matching image is 3.11, not near zero. Conditioning, not tensor identity, selects the image, and the ported flow-matching core is functionally correct. Trained for real, two Stage-1 seeds then run to ~15,800 steps and land within seed noise of each other at a held-out velocity-MSE of 0.17544 and 0.17547 — well below every short screen — and their checkpoints are the generations shown throughout.
The Bakeoffs: Muon Wins, Audited
With the pipeline validated, the stage ran two-arm bakeoffs, each with a pre-registered gate on held-out velocity-MSE. The optimiser verdict fired: Muon-hybrid beats all-AdamW at equal wall-clock, 0.18333 against 0.18609, and AdamW lost despite completing about 8% more steps, so its per-step disadvantage exceeds its throughput edge. This matches the sibling language-model program's Muon finding — the port's hot-learning-rate heritage transfers to flow-matching DiTs.
The audit bounds the claim precisely. The comparison is not a pure optimiser A/B: the Muon trunk matrices carried weight-decay 0.0 while all-AdamW carried 0.1, a regularisation confound on about 85% of the parameters, so this is a Muon-hybrid recipe against an all-AdamW recipe, not an isolated optimiser swap. The '9.2x seed noise' multiplier is withdrawn as a significance statement at n=1 per arm; what stands is directional — Muon led at every checkpoint and both AdamW arms landed above all four Muon runs. The learning-rate sweep was nearly flat, and the verdict is reported under the letter of the pre-registration: the rule required a gap no larger than the between-seed gap, and 0.00042 exceeded the 0.00030 seed-noise bound, so muon-lr 0.01 is reported as the screening winner.
A Perceptual Metric for What the Loss Cannot See
The program's only gating metric, held-out velocity-MSE, is insensitive to text content: on the Stage-1 checkpoint it reads about 0.2025 with each clip's real caption, 0.2025 with shuffled captions, and 0.2026 with no caption at all, and it moved only ~0.002 across an 8x parameter sweep. A metric that reads the same for a real and a shuffled caption cannot certify that an intervention changed the image, so we built a paired CLIP measure that can.
The consequence, measured: every gallery grid is the same 16 held-out prompts at the same seed, so a cell of two grids is two models' answer to one prompt. Encoding each cell with CLIP ViT-L/14 and comparing matched cells against a prompt floor (different prompts, same model: 0.7330) and a conservative seed floor (same recipe, different seed: ~0.85), three of six arm pairs — including the entire learning-rate sweep — are as similar to each other as two runs of the same recipe with a different seed. Those interventions moved the loss without moving the picture. Two headline results follow: the Muon-vs-AdamW outcome, Muon ahead at every checkpoint on loss, is a perceptual near-tie (0.8463, CI straddling the floor); and quality-injection, the one intervention that demonstrably changed the picture (0.7835, below every seed pair), scored lower on the primary metric. The finding is clean — a loss delta is not a picture delta — and the program measures both.
Guidance
An adversarial audit found that the trainer spent 10% of every batch training an unconditional branch the sampler never used: every image this program ever published was rendered at classifier-free guidance w=1.0 — that is, with guidance effectively off. Fixing it costs zero training GPU-hours; only the sampler changes. We ran a sweep on the same checkpoint and the same 16 held-out prompts, scoring the right criterion — whether the picture depends on its prompt, as a paired matched-minus-mismatched CLIPScore gap.
Three conclusions. Turning guidance on is well supported: the w=1 and w=5 confidence intervals do not overlap, and at w=1 the interval includes values indistinguishable from ignoring the prompt. On the point estimate w=3 is the optimum, not the w=5 that shipped — though the intervals from w=2 to w=7.5 overlap heavily, a plateau that 16 prompts cannot resolve, so the shipped value is arbitrary within a flat region rather than wrong. And the measured detail: guidance is primarily an aesthetic control — from w=1 to 7.5 saturation rises 70% while sharpness stays flat and the subject does not change; the car is already a car at w=1. The point stands confidently: guidance amplifies the scene the prompt already picked; it does not make the prompt pick a better subject.

Conditioning, Measured: a Breakout to 24% of the Ceiling
The loss could not see conditioning, so we built a metric that can and watched it over training. For each generated image we take its CLIPScore against its own caption minus its mean CLIPScore against the other 15 — a paired prompt-alignment gap that asks the one question worth asking, does the image depend on its prompt, with a bootstrap CI and a permutation p. The same 16 held-out prompts, seed, guidance and steps are used at every point, so training is the only variable, and the ground-truth images scored through the identical path set the ceiling.
The curve is the clearest positive result in the paper. At step 1,984 the gap is -0.0035 with p=0.78: the model is ignoring its prompt, indistinguishable from chance. It then holds a genuine ~14,000-step plateau at ~16% of the ceiling, and then it breaks out — by step 31,520 the gap is +0.0483, CI95 [+0.0347, +0.0630], p<=0.001, which is 24.0% of the ground-truth ceiling. This is a from-scratch model learning, measurably and significantly, to make its picture depend on its caption. Two honest bounds travel with it: the ceiling is low in absolute terms because the captions are raw CC12M alt-text (SEO strings), so 24% of that ceiling is a real signal, not a quality claim; and a coarser independent probe agrees — on 64 held-out prompts, a sample's nearest ground-truth-in-latent is its own caption's image 4 of 64 times (6.2%) against 1.6% chance (p=0.018).
Every interim claim is held to its receipt: the optimiser recipe is reported as provisionally selected, the LR verdict restated under the pre-registration, the w=5.0 ship corrected to w=3.0, and the arm-similarity screen's power disclosed. What's next is a roadmap: close out M1's gate with the judge-and-metric scorecard against the SD1.5-class baseline; power up the optimiser verdict with a weight-decay-matched Muon variant and more seeds; and push conditioning past 24% of ceiling with better captions than raw alt-text, more samples, and a stronger text-feature choice than the last hidden state. M1 is a confident result: the pipeline works, the model learns to follow its prompt, and we built the perceptual metric that proves it — which is exactly the knowledge the video rungs of the program are built on: demand both the loss and the picture, and never confuse one for the other.

Ground Truth, for Scale
The matched ground-truth images for the same 16 held-out prompts make the gap concrete. Reading the two grids together shows exactly what the model has and has not earned at this stage — coarse category and layout, yes; fine structure and legible text, not yet — and keeps the 24%-of-ceiling conditioning result honest about the quality it does and does not imply.

Abstract
We trained BlackMind's first text-to-image model, M1, a latent diffusion transformer from random initialisation on CC12M, on two consumer GPUs, with no downloaded generator weights, no LoRA, and no pretrained image prior; the only frozen components are a VAE and a text encoder used strictly as instruments. It is the pipeline-validation rung of a staged video-model program. The port is correct — a DiT-S memorises 64 fixed images and routes each caption to its own image 64 of 64 times from pure noise — and trained for real, two seeds reach a held-out velocity-MSE of 0.1754 over ~15.8k steps while the model draws recognisable cars, kitchens, food and garments from held-out prompts. The headline positive result is measured conditioning emergence: on a controlled CLIP prompt-alignment probe the model climbs from ignoring its caption at step 2k (p=0.78), through a 14,000-step plateau at ~16% of the ground-truth caption ceiling, to a breakout at 24% of ceiling by step 31.5k (paired gap +0.0483, CI95 [+0.0347,+0.0630], p<=0.001). The methodological contribution underneath it is a purpose-built perceptual metric. Because held-out velocity-MSE is insensitive to text content — it reads about 0.2025 with real, shuffled, and absent captions — we built a paired CLIP prompt-alignment measure that tracks conditioning directly, and used it to establish which training interventions change the picture and which only move the loss, a distinction only a perceptual metric can draw. A Muon-hybrid optimiser leads on loss (0.18333 vs 0.18609), a full classifier-free-guidance sweep sets w=3.0 as the operating point, and every interim claim is held to its on-disk receipt under a pre-registered audit.
More figures

Fig. 1The final checkpoint on its own: 16 held-out prompts (never trained on), sampled from pure noise at step 31,520. Recognisable cars, kitchens, food, garments and flowers with structural distortion and garbled text — the honest state of the image model.
