Skip to content

Transform your 3D texturing workflow with the power of generative AI, directly within Blender!

turn qwen 2.1 native rgba output off by default

sakalond3af2b7fa63dd
diff --git a/stablegen/core/properties.py b/stablegen/core/properties.py
index 9c9fcb7..00ac515 100644
--- a/stablegen/core/properties.py
+++ b/stablegen/core/properties.py
@@ -429,11 +429,13 @@ def register_properties(update_model_list, ControlNetUnit, LoRAUnit,
bpy.types.Scene.qwen_image_21_native_rgba = bpy.props.BoolProperty(
name="Native RGBA Output",
description=(
- "Wrap Qwen-Image 2.1 prompts with the native transparency instruction. "
- "When enabled, transparent generated pixels are excluded from texture "
- "projection. Disable it to let the prompt and model choose the background"
+ "Wrap Qwen-Image 2.1 prompts with the native transparency instruction, "
+ "so the model cuts the subject out on a transparent background. Off by "
+ "default: the model also cuts away surfaces it should paint, such as "
+ "ground and walls, which leaves them for worse-angled views. "
+ "Transparent generated pixels are never projected"
),
- default=True,
+ default=False,
update=update_parameters,
)
bpy.types.Scene.moss_depth_control = bpy.props.BoolProperty(
diff --git a/stablegen/ui/presets.py b/stablegen/ui/presets.py
index 80d42bf..4eb5573 100644
--- a/stablegen/ui/presets.py
+++ b/stablegen/ui/presets.py
@@ -89,16 +89,14 @@ PRESETS["QWEN 2.1 PRECISE"] = {
"DPM++ SDE and SGM Uniform with no Lightning LoRA. This combination "
"produced broad, coherent panels that followed the model silhouette and "
"continued well into later views. Best when adjacent cameras overlap "
- "well; sparse coverage can still introduce artifacts. Qwen 2.1 is also "
- "prompted to generate native RGBA views, independently of the sampler, "
- "and transparent pixels are excluded from texture projection."
+ "well; sparse coverage can still introduce artifacts."
),
"model_architecture": "qwen_image_21",
"steps": 20,
"cfg": 1.0,
"sampler": "dpmpp_sde",
"scheduler": "sgm_uniform",
- "qwen_image_21_native_rgba": True,
+ "qwen_image_21_native_rgba": False,
"qwen_guidance_background_color": mathutils.Color((0.5, 0.5, 0.5)),
"lora_units": [],
}
@@ -115,8 +113,7 @@ PRESETS["QWEN 2.1 PRECISE+"] = {
"Uses a 0.50 context trust threshold, a 3 px silhouette margin and "
"backface projection. Retains Qwen 2.1 Precise's projected-context "
"strategy, 20 steps, CFG 1.0, DPM++ SDE, SGM Uniform, blending exponent "
- "7 and native RGBA output, with no Lightning LoRA. Transparent pixels "
- "are excluded from projection. Research options are off. Best suited "
+ "7, with no Lightning LoRA. Research options are off. Best suited "
"to overlapping camera views; fine patterns and material continuity "
"can still vary, and standard Precise can work better on some assets."
),
@@ -141,8 +138,6 @@ PRESETS["QWEN 2.1 SAFE"] = {
"and details around the object. Uses the tested projection-oriented "
"configuration of 20 steps, CFG 1, DPM++ SDE and SGM Uniform with no "
"Lightning LoRA. It favors broad, geometry-following texture regions. "
- "Qwen 2.1 is prompted to generate native RGBA views independently of the "
- "sampler, and transparent pixels are excluded from texture projection. "
"It is slower when the initial reference is tall or large."
),
"model_architecture": "qwen_image_21",
@@ -150,7 +145,7 @@ PRESETS["QWEN 2.1 SAFE"] = {
"cfg": 1.0,
"sampler": "dpmpp_sde",
"scheduler": "sgm_uniform",
- "qwen_image_21_native_rgba": True,
+ "qwen_image_21_native_rgba": False,
"sequential_ipadapter": True,
"sequential_ipadapter_mode": "first",
"qwen_context_render_mode": "NONE",
@@ -168,8 +163,7 @@ PRESETS["QWEN 2.1 CONTEXT + REFERENCE"] = {
"depth guide for geometry and the initial view for overall appearance. "
"A 50-degree discard angle excludes grazing projections during generation, "
"then resets to 90 degrees for the final material; blending exponent is 7. "
- "Uses 20 steps, CFG 1.0, DPM++ SDE, SGM Uniform, native RGBA output and no "
- "LoRA. Transparent generated pixels are excluded from projection. "
+ "Uses 20 steps, CFG 1.0, DPM++ SDE, SGM Uniform and no LoRA. "
"Confidence-Gated Context, Acquisition Quality Weighting and Depth/Part "
"Projection Guard are off; legacy fallback dilation is 1 px, cleanup is "
"off and backface projection is allowed. Captures the current manually "
@@ -208,9 +202,8 @@ PRESETS["LOCAL EDIT (QWEN 2.1)"] = {
"Preserves the original Local Edit camera coverage and edge-feathering "
"settings so the edited projections blend into the existing material. "
"Uses 20 steps, CFG 1.0, DPM++ SDE, SGM Uniform, blending exponent 7 "
- "and native RGBA output, with no Lightning LoRA. Transparent generated "
- "pixels do not contribute to projection. This edits camera-selected "
- "areas; use Masked Edit generation mode for painted-mask selection."
+ "and no Lightning LoRA. This edits camera-selected areas; use Masked "
+ "Edit generation mode for painted-mask selection."
),
"architecture_mode": "qwen_image_21",
"model_architecture": "qwen_image_21",
@@ -220,7 +213,7 @@ PRESETS["LOCAL EDIT (QWEN 2.1)"] = {
"cfg": 1.0,
"sampler": "dpmpp_sde",
"scheduler": "sgm_uniform",
- "qwen_image_21_native_rgba": True,
+ "qwen_image_21_native_rgba": False,
"qwen_guidance_background_color": mathutils.Color((0.5, 0.5, 0.5)),
"lora_units": [],
}
@@ -233,10 +226,9 @@ PRESETS["REFINE (QWEN 2.1)"] = {
"as its editing input, making this suitable for overall color, art "
"style or surface-appearance changes. Preserves the original Qwen "
"Refine workflow settings. Uses 20 steps, CFG 1.0, DPM++ SDE, SGM "
- "Uniform, blending exponent 7 and native RGBA output, with no Lightning "
- "LoRA. Transparent generated pixels are excluded from projection. "
- "Use Local Edit for camera-targeted changes or Masked Edit generation "
- "mode to select an area by painting."
+ "Uniform, blending exponent 7 and no Lightning LoRA. Use Local Edit for "
+ "camera-targeted changes or Masked Edit generation mode to select an "
+ "area by painting."
),
"architecture_mode": "qwen_image_21",
"model_architecture": "qwen_image_21",
@@ -246,7 +238,7 @@ PRESETS["REFINE (QWEN 2.1)"] = {
"cfg": 1.0,
"sampler": "dpmpp_sde",
"scheduler": "sgm_uniform",
- "qwen_image_21_native_rgba": True,
+ "qwen_image_21_native_rgba": False,
"qwen_guidance_background_color": mathutils.Color((0.5, 0.5, 0.5)),
"lora_units": [],
}
@@ -268,8 +260,7 @@ for _base_name, _purpose in _QWEN_21_TURBO_PURPOSES.items():
"description": (
f"Qwen-Image 2.1 Turbo {_purpose}. Matches {_base_name.title()} except "
"for sampling: the Turbo model's trained 8-step schedule at CFG 1 with "
- "Euler, several times faster than 20 DPM++ SDE steps. Native RGBA "
- "output is on and transparent pixels are excluded from projection."
+ "Euler, several times faster than 20 DPM++ SDE steps."
),
"architecture_mode": "qwen_image_21_turbo",
"model_architecture": "qwen_image_21_turbo",