refactor(architectures): drive latent-space facts from the registry

invokeai/app/invocations/denoise_latents.py

@@ -38,6 +38,7 @@
 from invokeai.app.invocations.t2i_adapter import T2IAdapterField
 from invokeai.app.services.shared.invocation_context import InvocationContext
 from invokeai.app.util.controlnet_utils import prepare_control_image
+from invokeai.backend.architectures import get_max_unet_downscale
 from invokeai.backend.ip_adapter.ip_adapter import IPAdapter
 from invokeai.backend.model_manager.configs.factory import AnyModelConfig
 from invokeai.backend.model_manager.taxonomy import BaseModelType, ModelVariantType
@@ -676,17 +677,13 @@ def run_t2i_adapters(
             t2i_adapter_model_config = context.models.get_config(t2i_adapter_field.t2i_adapter_model.key)
             image = context.images.get_pil(t2i_adapter_field.image.image_name, mode="RGB")
 
-            # The max_unet_downscale is the maximum amount that the UNet model downscales the latent image internally.
-            if t2i_adapter_model_config.base == BaseModelType.StableDiffusion1:
-                max_unet_downscale = 8
-            elif t2i_adapter_model_config.base == BaseModelType.StableDiffusionXL:
-                max_unet_downscale = 4
+            # Raises for a base without a UNet, before the BGR swap below -- same order as before.
+            max_unet_downscale = get_max_unet_downscale(t2i_adapter_model_config.base)
 
+            if t2i_adapter_model_config.base == BaseModelType.StableDiffusionXL:
                 # SDXL adapters are trained on cv2's BGR outputs
                 r, g, b = image.split()
                 image = Image.merge("RGB", (b, g, r))
-            else:
-                raise ValueError(f"Unexpected T2I-Adapter base model type: '{t2i_adapter_model_config.base}'.")
 
             t2i_adapter_model: T2IAdapter
             with context.models.load(t2i_adapter_field.t2i_adapter_model) as t2i_adapter_model:
@@ -985,6 +982,11 @@ def step_callback(state: PipelineIntermediateState) -> None:
             #    ext = extension_field.to_extension(exit_stack, context, ext_manager)
             #    ext_manager.add_extension(ext)
             self.parse_controlnet_field(exit_stack, context, self.control, ext_manager)
+            # NOTE: this decides the BGR swap from the *UNet's* base, while run_t2i_adapters above
+            # decides it from each *adapter's* base. The two disagree for an SD1 adapter on an SDXL
+            # UNet. Left as-is deliberately: fixing it changes behaviour, and this refactor does
+            # not. The natural fix is to fold `bgr_input` into UNetDownscaleFacet so both paths read
+            # one declaration.
             bgr_mode = self.unet.unet.base == BaseModelType.StableDiffusionXL
             self.parse_t2i_adapter_field(exit_stack, context, self.t2i_adapter, ext_manager, bgr_mode)
 

invokeai/app/invocations/ideogram4_denoise.py

@@ -12,16 +12,13 @@
 from invokeai.app.invocations.model import TransformerField
 from invokeai.app.invocations.primitives import LatentsOutput
 from invokeai.app.services.shared.invocation_context import InvocationContext
-from invokeai.app.util.step_callback import (
-    FLUX2_LATENT_RGB_BIAS,
-    FLUX2_LATENT_RGB_FACTORS,
-    sample_to_lowres_estimated_image,
-)
+from invokeai.backend.architectures import get_latent_space
 from invokeai.backend.ideogram4 import run_ideogram4_denoise
 from invokeai.backend.ideogram4.latent_norm import get_latent_norm
 from invokeai.backend.ideogram4.sampler_configs import PRESETS
 from invokeai.backend.ideogram4.sampling_utils import unpatchify_and_denormalize
 from invokeai.backend.ideogram4.transformer_pair import Ideogram4TransformerPair
+from invokeai.backend.model_manager.taxonomy import BaseModelType
 from invokeai.backend.stable_diffusion.diffusion.conditioning_data import Ideogram4ConditioningInfo
 from invokeai.backend.util.devices import TorchDevice
 
@@ -122,12 +119,13 @@ def invoke(self, context: InvocationContext) -> LatentsOutput:
         assert isinstance(info, Ideogram4ConditioningInfo)
         llm_features = info.prompt_embeds.to(device=device, dtype=torch.float32)
 
-        # Progress-preview setup: Ideogram uses a FLUX.2-style 32-channel VAE, so the FLUX.2
-        # latent->RGB factors give a usable (approximate) low-res preview of the forming image at each
-        # step, without a full VAE decode. Denormalization params come from get_latent_norm (no VAE).
+        # Progress-preview setup: Ideogram's latent space gives a usable (approximate) low-res
+        # preview of the forming image at each step, without a full VAE decode. Denormalization
+        # params come from get_latent_norm (no VAE). This does not go through
+        # diffusion_step_callback: the callback signature here is (step, total, packed_latents) and
+        # the latents must be unpatchified and denormalized first.
         latent_shift, latent_scale = get_latent_norm()
-        rgb_factors = torch.tensor(FLUX2_LATENT_RGB_FACTORS, dtype=torch.float32)
-        rgb_bias = torch.tensor(FLUX2_LATENT_RGB_BIAS, dtype=torch.float32)
+        latent_space = get_latent_space(BaseModelType.Ideogram4)
 
         def step_callback(step: int, total: int, packed_latents: torch.Tensor) -> None:
             preview = None
@@ -138,11 +136,7 @@ def step_callback(step: int, total: int, packed_latents: torch.Tensor) -> None:
                     latent_shift.to(packed_latents.device),
                     latent_scale.to(packed_latents.device),
                 )
-                preview = sample_to_lowres_estimated_image(
-                    samples=vae_latent,
-                    latent_rgb_factors=rgb_factors.to(vae_latent.device),
-                    latent_rgb_bias=rgb_bias.to(vae_latent.device),
-                )
+                preview = latent_space.preview(vae_latent)
             except Exception:
                 # A preview must never break generation — fall back to a plain progress signal.
                 preview = None
@@ -151,7 +145,10 @@ def step_callback(step: int, total: int, packed_latents: torch.Tensor) -> None:
                     "Running Ideogram 4 denoising",
                     step / total,
                     preview,
-                    (preview.width * 8, preview.height * 8),
+                    (
+                        preview.width * latent_space.spatial_compression,
+                        preview.height * latent_space.spatial_compression,
+                    ),
                 )
             else:
                 context.util.signal_progress("Running Ideogram 4 denoising", step / total)