PRXPixel

Photoroom’s 7B pixel-space variant of PRX. There is no VAE: the transformer denoises raw RGB in 16×16 pixel patches, predicts the clean image rather than a velocity, and starts from noise with std 2.0. Text comes from a Qwen3-VL text tower.

weights: Photoroom/prxpixel-t2i ↗
org: Photoroom
modality: image
tasks: text-to-image
license: Apache 2.0
native output: 1024×1024
total params: 8.72B
model.arch: prx_pixel

Components

role	model	params	size	dtype	trained
Transformer	PRX-7B pixel DiT Photoroom/prxpixel-t2i/transformer PRXTransformer2DModel The diffusers class is still an unmerged PR, so ai-toolkit vendors its own copy of the transformer and a small preview sampler.	7.00B 7,003,535,872	14.01 GB	bf16	yes
Text encoder	Qwen3-VL text tower (28 layers, 2048-d) Photoroom/prxpixel-t2i/text_encoder transformers.Qwen3VLTextModel Text model only, no vision tower. Last hidden state.	1.72B 1,720,574,976	3.44 GB	bf16	no
Tokenizer	Qwen2 tokenizer Photoroom/prxpixel-t2i/tokenizer transformers.Qwen2TokenizerFast	—	—	—	—
total		8.72B	17.45 GB

Latent space

spatial

1×

channels

patch

16×16

pixels per token: 16×16
notes: Pixel space: no VAE. The "latent" is the RGB image in [-1, 1]; ai-toolkit uses an identity FakeVAE so encode and decode do nothing. Each 16×16×3 patch (768 values) goes through a two-layer bottleneck projection (768 → 768 → 3584) into one token.

input	latent (c×t×h×w)	tokens
1024×1024	3×1×1024×1024	4,096	native, ai-toolkit sample default
512×512	3×1×512×512	1,024

Architecture

Blocks: 24
Hidden size: 3584 (28 heads × 128)
FFN size: 12544 (mlp_ratio 3.5)
Text conditioning: Image queries attend to text + image keys/values in every block. Text states (2048-d, 256 tokens) are projected but not updated
Resolution conditioning: The output resolution is embedded into the timestep modulation
Objective: Flow matching with x-prediction (the model outputs the clean image), shift 3.0, noise std 2.0
Norm / position: QK norm, 2D RoPE on image tokens (axes 64/64)

In AI Toolkit

model.arch: prx_pixel
UI label: PRXPixel (pixel space) (image)
model.name_or_path: Photoroom/prxpixel-t2i
source: extensions_built_in/diffusion_models/prx_pixel_t2i/prx_pixel_t2i.py
extensions_built_in/diffusion_models/prx_pixel_t2i/src/transformer_prx.py
extensions_built_in/diffusion_models/prx_pixel_t2i/src/pipeline.py
toolkit/models/v2/text_encoders/qwen3_vl.py
toolkit/models/FakeVAE.py
extensions_built_in/diffusion_models/ui.tsx
extra UI sections: model.low_vram, model.layer_offloading

UI defaults

quantize / quantize_te: true / true
low_vram: true
timestep_type: linear
network.conv: disabled (linear LoRA only)

Specifics

Resolution: Buckets and sample sizes snap to multiples of 16 (the patch size; there is no VAE).
Loss: x-prediction: the model output is compared with the clean image, not with noise − clean. The x0 → velocity conversion only happens while sampling.
Noise: Training noise and the starting noise for samples are randn × 2.0. noise_offset is applied on top.
Prompt encoding: Prompts are padded or truncated to exactly 256 tokens, with an attention mask.
Sampling: Flowmatch Euler, shift 3.0. CFG is applied on the x0 prediction, then converted to velocity with t clamped at 0.05. The vendor example uses 28 steps at guidance 5.0.
LoRA target: PRXTransformer2DModel
Saving: LoRAs use the ComfyUI key prefix. Full fine-tunes save transformer/ in Diffusers format.
Metadata base version: prx_pixel

Example config

not verified

job: extensionconfig:  name: "my_prx_pixel_lora_v1"  process:    - type: "sd_trainer"      training_folder: "output"      device: cuda:0      network:        type: "lora"        linear: 32        linear_alpha: 32      save:        dtype: float16        save_every: 250        max_step_saves_to_keep: 4      datasets:        - folder_path: "/path/to/images"          caption_ext: "txt"          caption_dropout_rate: 0.05          resolution: [512, 768, 1024]      train:        batch_size: 1        steps: 2000        gradient_checkpointing: true        noise_scheduler: "flowmatch"        timestep_type: "linear"        optimizer: "adamw8bit"        lr: 1e-4        dtype: bf16      model:        name_or_path: "Photoroom/prxpixel-t2i"        arch: "prx_pixel"        quantize: true        quantize_te: true        low_vram: true      sample:        sampler: "flowmatch"        sample_every: 250        width: 1024        height: 1024        guidance_scale: 5.0        sample_steps: 28        prompts:          - "A front-facing portrait of a lion in the golden savanna at sunset."