Architecture Studio

SD 2.1. Mari

Inspect the structural blueprint separately from run history. Model-level diagrams and run-linked exports stay together here without crowding the main model page.

Backbone SD 2.1

Architecture Files 0

Latest Update 9d ago

Available Renders

Only an ASCII representation is available for this model.

ASCII Architecture

text rendering

SD2JointModel [nn.Module]

config:
  - prediction_type <- prediction_type = "v_prediction"
  - use_ema <- use_ema = True
  - freeze_encoder <- freeze_encoder = False
  - lambda_dehaze <- lambda_dehaze = 1.0
  - lambda_depth <- lambda_depth = 1.0
  - use_visibility_head <- use_visibility_head = False
  - visibility_hidden_channels <- visibility_hidden_channels = 32
  - use_cross_task_fusion <- use_cross_task_fusion = False
  ... +16 more

submodules:
  - vae (AutoencoderKL.from_pretrained)
  - noise_scheduler (DDPMScheduler.from_pretrained)
  - inference_scheduler (DDIMScheduler.from_pretrained)
  - conv_in (unet.conv_in)
  - time_proj (unet.time_proj)
  - time_embedding (unet.time_embedding)
  - down_blocks (unet.down_blocks)
  - mid_block (unet.mid_block)
  - dehaze_up_blocks (unet.up_blocks)
  - dehaze_conv_norm_out (unet.conv_norm_out)
  ... +10 more

methods:
  - forward(hazy_images, clear_images, depth_images, valid_mask, timesteps, visibility_loss_scale) -> encode_images, encode_images, encode_images
  - forward_dehaze(noisy_clear, hazy_latents, timesteps, encoder_hidden_states, vis_ctx)
  - forward_depth(noisy_depth, hazy_latents, timesteps, encoder_hidden_states, vis_ctx) -> _time_embed, _encoder_forward, _collect_dehaze_guidance
  - _decoder_blocks_forward(sample, down_block_res_samples, emb, encoder_hidden_states, up_blocks, up_adapters, vis_ctx, collect_indices, stop_after_block, fusion_features, fusion_adapters)
  - _decoder_forward(sample, down_block_res_samples, emb, encoder_hidden_states, up_blocks, conv_norm_out, conv_out, up_adapters, out_adapter, vis_ctx, fusion_features, fusion_adapters) -> _decoder_blocks_forward
  - _encoder_forward(sample, emb, encoder_hidden_states)
  - _task_forward(noisy_target, hazy_latents, timesteps, encoder_hidden_states, up_blocks, conv_norm_out, conv_out, up_adapters, out_adapter, vis_ctx) -> _time_embed, _encoder_forward
  - _collect_dehaze_guidance(sample, down_block_res_samples, emb, encoder_hidden_states, vis_ctx) -> _decoder_blocks_forward
  - _predict_clean_latents(noisy_latents, model_pred, timesteps)
  - _sample_training_noise(latents, timesteps, task)
  ... +2 more

Inspector

summary

Architecture files are not available, so there is no file-level metadata to inspect.

Model Flow

Deterministic Runtime View

Building the deterministic flow graph from the model source in the background.