Diffusion, flow matching and the samplers that set your cost per image.
25 traps, Jun 2026 to Jul 2026. Complete.
Set in interviews at Midjourney (13), OpenAI (5), Google DeepMind (3) and Anthropic (1).
Each trap: the interviewer’s question, the answer most candidates give, and the mechanism that breaks it. The full answers are on Substack.
Generative AI Research Engineer interview at Black Forest Labs
“Your first reflow pass straightened your sampling trajectories and let you drop steps. So you ran three more reflow iterations chasing single-step generation, and your FID quietly got worse. Why?”
Generative Vision #25
Jul 4, 2026
Research Engineer interview at Google DeepMind
“Your 1000-step diffusion teacher is too slow for production, so you distill it into a single-step student. The output is garbage. You swap in a bigger student. Still garbage. What’s actually broken?”
Generative Vision #24
Jul 3, 2026
Senior ML Engineer interview at Midjourney
“Your text-to-image inference costs are killing the business. Your first instinct is to shrink the student model. Walk me through why that’s the wrong move, and what you’d cut instead.”
Generative Vision #23
Jul 2, 2026
Generative AI Engineer interview at Midjourney
“You DreamBooth a model on 5 photos of a client’s product. It generates that product perfectly, but now everything else it draws looks broken. Why, and how do you fix it without re-collecting data?”
Generative Vision #22
Jun 30, 2026
Senior ML Engineer interview at Midjourney
“Your text-to-image model crushes it on internal evals, but users complain the outputs look flat. They type ‘a teddy bear reading a book’ and get garbage. The weights are fine. What’s actually broken?”
Generative Vision #21
Jun 29, 2026
Machine Learning Engineer interview at OpenAI
“You fine-tuned your image model against a reward model. Your alignment scores jumped 30%. But humans say the outputs got worse. What happened and how do you stop it?”
Generative Vision #20
Jun 28, 2026
Senior ML Engineer interview at Midjourney
“Your text-to-image model renders the prompt correctly, right objects, right layout, but every output looks flat and amateur. Walk me through your fix.”
Generative Vision #19
Jun 27, 2026
Senior ML Engineer interview at OpenAI
“You added a REPA loss to your diffusion transformer to speed up training. It barely helped. What did you get wrong?”
Generative Vision #18
Jun 26, 2026
Senior ML Engineer interview at Midjourney
“You trained a diffusion model that’s gorgeous at 256px. You scale it to 1024px, keep the exact same noise schedule, and the outputs get quietly worse. Same architecture, same loss, same schedule. What broke?”
Generative Vision #17
Jun 25, 2026
Senior ML Engineer interview at Midjourney
“Your DiT trains beautifully at 512×512. You bump inference to 1024×1024 and it generates garbage, warped anatomy, repeated limbs, a teddy bear with three faces. Before you touch the VAE or the sampler, where do you look first?”
Generative Vision #16
Jun 24, 2026
Senior ML Engineer interview at Midjourney
“Your DiT trains beautifully at 512×512. You bump inference to 1024×1024 and it generates garbage, warped anatomy, repeated limbs, a teddy bear with three faces. Before you touch the VAE or the sampler, where do you look first?”
Generative Vision #15
Jun 23, 2026
Senior AI Engineer interview at Stability AI
“You’re building an MMDiT. Your teammate wants to share the feed-forward weights across text and image tokens, single-stream, because it’s simpler and cheaper. What are you actually giving up?”
Generative Vision #14
Jun 22, 2026
Senior AI Engineer interview at Midjourney
“You switched your text-to-image model from cross-attention to joint attention. Walk me through what actually changes about how text and image tokens talk to each other, and what specific failure that fixes.”
Generative Vision #13
Jun 21, 2026
Senior AI Engineer interview at OpenAI
“You doubled your DiT’s parameter count expecting FID to drop. It barely moved. What did you miss?”
Generative Vision #12
Jun 20, 2026
Senior AI Engineer interview at OpenAI
“You’re building a diffusion transformer. How do you inject the timestep and class condition into each block, and why?”
Generative Vision #11
Jun 19, 2026
Senior ML Engineer interview at Midjourney
“Your teammate wants to ship a U-Net for your new high-res image model because ‘convolutions capture both local and global features.’ You disagree. Defend it.”
Generative Vision #10
Jun 18, 2026
Senior AI Engineer interview at Midjourney
“Your text-to-image model nails single subjects, but prompt it with ‘a brown teddy bear next to a white wall’ and the colors bleed across the entire frame. An intern on your team says ‘add more training data.’ Why is that wrong, and where is this actually breaking?”
Generative Vision #9
Jun 17, 2026
Senior GenAI Engineer interview at Google DeepMind
“How does Conditional Flow Matching bypass the intractability of learning the marginal vector field?”
Generative Vision #8
Jun 16, 2026
Senior AI Engineer interview at Google DeepMind
“In continuous generative models, what is the exact physical difference between a ‘score’ and a ‘vector field’?”
Generative Vision #7
Jun 15, 2026
Senior AI Engineer interview at Anthropic
“How does the formulation of time and noise in Flow Matching fundamentally differ from standard DDPMs?”
Generative Vision #6
Jun 14, 2026
Senior AI Engineer interview at Midjourney
“Your score estimator loss is perfectly converged after 400 hours on an A100 cluster. But during inference, deterministically following the gradient generates the exact same 3 hyper-average images on repeat. Why?”
Generative Vision #5
Jun 12, 2026
Senior AI Engineer interview at Midjourney
“Images are discrete RGB values from 0 to 255. Diffusion math assumes we sample from a standard normal distribution. If a data engineer feeds raw 0-255 pixel tensors directly into the training pipeline without continuous float scaling, how does this mathematically break the variance-preserving nature of the forward process?”
Generative Vision #4
Jun 11, 2026
Senior AI Engineer interview at OpenAI
“Calculating the exact marginals for a 1,000-step diffusion trajectory is computationally impossible. Yet, the DDPM loss collapses into a simple L2 regression. What specific structural assumption saves us from infinite computational complexity?”
Generative Vision #3
Jun 10, 2026
Senior AI Engineer interview
“We are training a diffusion model and need to fetch the noisy state at step 500. Code it.”
Generative Vision #2
Jun 9, 2026
Senior AI Engineer interview at Midjourney
“Your diffusion model generates photorealistic textures, but the global shapes are completely mangled ( for example three-headed teddy bears ). The architecture is flawless. What phase of your forward noise schedule (𝛽_𝑡) is failing, and why?”
Generative Vision #1
Jun 8, 2026