MCPcopy Create free account
hub / github.com/tdrussell/diffusion-pipe / forward

Method forward

models/flux2.py:410–425  ·  view source on GitHub ↗
(self, inputs)

Source from the content-addressed store, hash-verified

408
409 @torch.autocast('cuda', dtype=AUTOCAST_DTYPE)
410 def forward(self, inputs):
411 img, pe, txt_len, img_len, hw, vec_orig, *vec = inputs
412
413 # Extract image portion (after text)
414 img = img[:, txt_len.item():, ...]
415 # Only process the noisy image portion (control image was concatenated after if present)
416 img = img[:, :img_len.item(), ...]
417
418 out = self.final_layer(img, vec_orig)
419 out = out[:, :img_len.item()]
420 h_orig = hw[0].item()
421 w_orig = hw[1].item()
422 patch_size = self.patch_size
423 h_len = ((h_orig + (patch_size // 2)) // patch_size)
424 w_len = ((w_orig + (patch_size // 2)) // patch_size)
425 return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=patch_size, pw=patch_size)[:,:,:h_orig,:w_orig]

Callers

nothing calls this directly

Calls

no outgoing calls

Tested by

no test coverage detected