--- title: Where role collapse happens in diffusion models date: 2026-10-23 summary: Prompt SDXL for a mouse chasing a cat and you usually get a cat chasing a mouse. My first suspect was CLIP reading prompts as a bag of words, but after ruling out the encoder, guidance, and cross-attention, the problem seems to sit in the UNet itself. --- If you prompt SDXL with "a mouse chasing a cat", about three times in four you get a cat chasing a mouse. The same goes for "a horse riding an astronaut" or "a doll holding a grandpa": both entities show up, in the roles they usually have. [Malakouti and Kovashka](https://arxiv.org/abs/2503.10037) call this role collapse. I wanted to know which part of the pipeline is responsible, so I went through the obvious suspects one at a time. My answer so far is that the text encoder isn't the problem: the denoiser seems to get the roles and draws the usual pairing anyway. The prompts follow their RoleBench setup: 10 relations, each in a common direction and a rare reversed one, active and passive voice, 4 images per prompt. A vision-language judge labeled each image correct, reversed, or neither. On a few hundred images I labeled myself, it sometimes calls a direction where I'd say neither (so the reversed rates are probably a bit high), but when it picks a direction it agrees with me. ## suspect one: CLIP This was my first guess. CLIP is known to treat captions a lot like a bag of words ([Yuksekgonul et al., 2023](https://arxiv.org/abs/2210.01936)), and it reads left to right, so in "a photo of one cat chasing one mouse" the cat token is computed before the verb exists. A linear probe for the role of the first noun sits at chance (0.50) in both of SDXL's CLIP encoders, so SDXL really is getting role-blind noun tokens. T5-XXL is the opposite. The first noun's role is decodable at about 1.00, including on relations the probe never saw, and in later layers the probe also works on sentences like "one mouse that one cat is chasing", which have no "being" in them. My read is that T5 encodes the actual role, not just word order. SD3.5-Large uses T5 alongside CLIP, which makes it a clean test. If CLIP were the cause, SD3.5 should do better. It doesn't: ``` rare prompts, reversed rate SDXL 0.74 SD3.5-Large 0.72-0.75 SD3.5, T5 only 0.45-0.57 ``` Zeroing the CLIP inputs lowers the reversed rate, but almost all of that moves to "neither" (a missing lion, a baby held instead of fed), and correct only goes up by about 4 points. When both entities are drawn, they still have the usual roles. ## suspect two: weak conditioning Maybe the model has the roles but the prior outweighs them, and stronger guidance would push them through. On SDXL, CFG 3, 7, and 12 all leave rare prompts 68-74% reversed, and so does passive voice (78%) and IterComp, a fine-tune trained for composition (76%). The pictures do change between settings, just not who is doing what. [[demo:role-collapse-rates]] ## suspect three: cross-attention Another option is that the noun tokens leak into each other's image regions. They don't seem to. In reversed images the attention maps for the two nouns are as separate as in correct ones (top-10% IoU 0.054 vs 0.053). Each noun gets its own region, and the model draws that entity in its usual pose. ## suspect four: some part of the conditioning Last, I tried splicing on SDXL: replace the verb, EOS, or pooled embedding with the one from the role-swapped prompt and see whether the drawn roles follow. None of these moved the reversed rate by more than about 4 points, which is within noise. My guess at why comes from comparing the two prompts with no splicing at all: ``` P(drawn as "A verbs B"), prompt "A verbs B" vs "B verbs A" common pairs 0.76 vs 0.74 other pairs 0.48 vs 0.45 ``` Swapping the whole sentence changes the result by about 3 points. Most of the outcome is set by which two nouns are in the prompt, so there isn't much role signal for a splice to find. ## what's left That leaves the denoiser itself (the UNet in SDXL, the MMDiT in SD3.5). With T5 the roles are in the text embedding, and SD3.5 draws what's typical for the noun pair anyway. If the failure really is inside the UNet, then a model that does draw reversed roles, like the LoRA from the ReBind paper, should differ from base SDXL somewhere in there. My plan is to train a [block-sparse featurizer](https://arxiv.org/abs/2606.25234) (BSF, a sparse autoencoder whose features are small subspaces) on UNet activations from both models, and look for a block that tracks the reversal. The part I'm most curious about is whether that block is separate from the ones for the subject and object. If it is, role would be its own feature in the UNet, something you could probably steer directly.