danielhanchen · Oct 21, 2026 · Oct 21, 2026 · Oct 21, 2026 · Oct 21, 2026 · Oct 21, 2026
diff --git a/.github/workflows/unsloth-sd-prebuilt.yml b/.github/workflows/unsloth-sd-prebuilt.yml
@@ -389,6 +389,16 @@ jobs:
         if: ${{ matrix.cpu_floor != '' }}
         run: bash tooling/scripts/unsloth/cpu_variants_gate.sh check "dist/sd-${{ needs.resolve.outputs.tag }}-bin-${{ matrix.label }}.zip" "${{ matrix.cpu_floor }}"
 
+      # The runner has the newest ISA the bundle uses, so it cannot see a variant library that crashes
+      # older CPUs while ggml dlopens every variant to score it. Replay startup on emulated CPUs: Intel SDE
+      # Nehalem .. Sapphire Rapids on x86_64, qemu Cortex-A53 .. max on aarch64.
+      - name: CPU dispatch gate (emulated CPUs)
+        if: ${{ matrix.cpu_floor != '' && runner.os == 'Linux' }}
+        run: |
+          set -euo pipefail
+          if [ "$(uname -m)" = "aarch64" ]; then sudo apt-get install -y -qq qemu-user; fi
+          bash tooling/scripts/unsloth/cpu_dispatch_gate.sh "dist/sd-${{ needs.resolve.outputs.tag }}-bin-${{ matrix.label }}.zip"
+
       - name: Upload bundle
         uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02  # v4.6.2
         with:

diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -328,6 +328,13 @@ if (NOT SD_USE_SYSTEM_GGML)
     add_definitions(-DGGML_MAX_NAME=160)
 endif()
 
+# x86 and aarch64 CPU: build llamafile's tinyBLAS GEMM (as llama.cpp does). ggml-patches/0017 picks the weight types
+# it takes per CPU from measurements (Q8_0 on x86; F16 / BF16 where tinyBLAS was faster, see ggml-cpu.c);
+# GGML_CPU_SGEMM_TYPES overrides the set.
+if (NOT DEFINED GGML_LLAMAFILE AND CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|AMD64|amd64|x64|aarch64|arm64|ARM64)$")
+    set(GGML_LLAMAFILE_DEFAULT ON)
+endif()
+
 # deps
 include(cmake/ggml.cmake)
 

diff --git a/README.md b/README.md
@@ -48,6 +48,7 @@ API and command-line option may change frequently.***
     - [FLUX.2-dev/FLUX.2-klein](./docs/flux2.md)
     - [FLUX.3 Image Klein / Klein Base](./docs/flux3.md)
     - [MOSS-Image-Base / MOSS-Image-SFT](./docs/moss_image.md)
+    - [Fibo 2.0 Dev](./docs/fibo2.md)
     - [Lens](./docs/lens.md)
     - [Chroma](./docs/chroma.md)
     - [Chroma1-Radiance](./docs/chroma_radiance.md)
@@ -79,6 +80,7 @@ API and command-line option may change frequently.***
     - [Boogu Image Edit](./docs/boogu_image.md)
     - [Mage-Flow-Edit](./docs/mage_flow.md#image-editing)
     - [LLaDA-Image Edit](./docs/llada_image.md#image-editing)
+    - [Fibo 2.0 Dev / Edit Turbo](./docs/fibo2.md#editing-and-turbo)
   - Video Models
     - [Wan2.1/Wan2.2](./docs/wan.md)
     - [MiniMax-H3](./docs/minimax_h3.md)

diff --git a/assets/fibo2/dev_edit.png b/assets/fibo2/dev_edit.png
diff --git a/assets/fibo2/t2i.png b/assets/fibo2/t2i.png
diff --git a/docs/adetailer.md b/docs/adetailer.md
@@ -1,9 +1,10 @@
 # ADetailer
 
 `sd-cli` can run a YOLOv8 object detector on an existing or newly generated
-image and perform a cropped inpaint pass for every detected object. The first
-implementation supports YOLOv8 detection checkpoints. YOLOv8 segmentation and
-MediaPipe models are not supported yet.
+image, and `sd-server` can run it after image generation. Both perform a cropped
+inpaint pass for every detected object. The implementation supports YOLOv8
+detection checkpoints. YOLOv8 segmentation and MediaPipe models are not
+supported yet.
 
 ## Convert a detector
 
@@ -68,6 +69,36 @@ ADetailer can also run automatically after normal image generation:
   --extra-ad-args "confidence=0.3,denoising_strength=0.4,inpaint_width=512,inpaint_height=512"
 ```
 
+The same options can be supplied when starting `sd-server`. When `--ad-model` is set, its detector runs after OpenAI-compatible and native asynchronous image generation. SDAPI requests enable ADetailer through `alwayson_scripts`, as described below. If no object is detected, the original image is returned unchanged. RGBA output keeps its original alpha channel, including when detected regions are repaired. Detector initialization or detail-pass failures are reported as request failures.
+
+Server image-generation requests limit base, hires, and detail-pass dimensions to 4096 pixels per side. ADetailer step overrides are limited to 1–100; `steps=0` in `extra_ad_args` keeps inheriting the main generation steps. These limits also apply to inherited dimensions and `extra_ad_args` overrides. Requests exceeding the limits return a parameter error before generation. CLI limits are unchanged.
+
+### SDAPI requests
+
+`/sdapi/v1/txt2img` and `/sdapi/v1/img2img` accept the WebUI ADetailer argument format:
+
+```json
+{
+  "prompt": "portrait photo",
+  "alwayson_scripts": {
+    "ADetailer": {
+      "args": [true, false, {
+        "ad_model": "face_yolov8n.pt",
+        "ad_prompt": "[PROMPT], detailed face",
+        "ad_use_steps": true,
+        "ad_steps": 20
+      }]
+    }
+  }
+}
+```
+
+The detector must be configured at server startup with `--ad-model`. `ad_model` must match that file's basename; the original `.pt` basename is also accepted for a converted detector. Requests cannot load arbitrary detector paths. Omitting the script, setting its first argument to `false`, selecting `"None"`, or setting `ad_tab_enable` to `false` disables detailing. Dictionary-only argument arrays are also accepted. Multiple enabled dictionaries run in order using the configured detector.
+
+Supported settings include prompts, confidence, mask filtering by area or confidence, mask count and area ratios, dilation/erosion, offsets, mask merging/inversion, blur, denoising strength, masked or full-image inpainting, padding, and optional overrides for inpaint dimensions, steps, CFG, sampler/scheduler, and CLIP skip. The corresponding `ad_use_*` flags control whether overrides apply. Otherwise dimensions, steps, CFG, sampler, scheduler, and CLIP skip inherit the main request. Denoising strength defaults to `0.4`. Pixel mask settings are limited to the larger output dimension, with a minimum limit of 32 and a maximum of 16384 pixels.
+
+Model/VAE switching, non-default noise multipliers, face restoration, ControlNet, and skipping the initial img2img pass are not supported; active requests for these features return an error. Unknown settings and invalid active values also return errors. During detailing, `/sdapi/v1/progress` remains active with an `adetailer` job; its percentage and preview stay fixed until detailing finishes.
+
 An empty ADetailer prompt inherits the main prompt. `[PROMPT]` inserts the main
 prompt, `[SEP]` assigns different prompts to consecutive masks, and `[SKIP]`
 skips the corresponding mask.
@@ -82,6 +113,7 @@ All settings other than the detector path and prompts are passed through
 | `nms` | `0.45` | NMS IoU threshold |
 | `max_detections` | `100` | Maximum detections retained after NMS |
 | `mask_k_largest` | `0` | Keep only the largest K masks; zero keeps all |
+| `mask_filter_method` | `area` | Rank retained masks by `area` or `confidence` |
 | `mask_min_ratio` | `0` | Minimum bbox area relative to the image |
 | `mask_max_ratio` | `1` | Maximum bbox area relative to the image |
 | `dilate_erode` | `4` | Positive values dilate; negative values erode |
@@ -90,6 +122,7 @@ All settings other than the detector path and prompts are passed through
 | `merge_masks`, `invert_mask` | `false` | Boolean alternatives to `mask_mode` |
 | `mask_blur` | `4` | Final composite feather radius |
 | `inpaint_padding` | `32` | Padding around the detected region |
+| `inpaint_only_masked` | `true` | Crop to the detected region; `false` uses the full image with the detection mask |
 | `inpaint_width`, `inpaint_height` | mode-specific | `512x512` after generation; input/output size in `adetailer` mode |
 | `denoising_strength` | mode-specific | `0.4` after generation; inherits `--strength` in `adetailer` mode |
 | `steps` | `0` | Detail steps; zero inherits the main generation |

diff --git a/docs/backend.md b/docs/backend.md
@@ -241,9 +241,13 @@ sd-cli -m model.safetensors -p "a cat" --conv-direct off,diffusion=on
   CUDA 2D requires a compiled Turing-or-newer MMA architecture; HIP 2D requires
   the supported CDNA MFMA or RDNA WMMA architectures described below for flash
   attention. MUSA 2D has a SIMT implementation. CUDA, HIP, and MUSA 3D also have
-  a SIMT implementation, so they do not require MMA. Other backends default to
-  `off`. Results and decision logs are cached per compute backend and dimension
-  in the context, as with flash attention. Each dimension is selected independently,
+  a SIMT implementation, so they do not require MMA. The CPU backend selects
+  `on` for 2D and 3D when ggml carries the row-run direct kernels
+  (`scripts/unsloth/ggml-patches/0014`, applied by the prebuilt workflow) and
+  `off` otherwise: the direct kernels use a fixed patch buffer instead of the
+  full im2col matrix and match im2col bit for bit with F16 weights. Other
+  backends default to `off`. Results and decision logs are cached per compute
+  backend and dimension in the context, as with flash attention. Each dimension is selected independently,
   and all compute devices in a module must qualify. Parameter-storage devices do
   not affect this choice.
 

diff --git a/docs/edit.md b/docs/edit.md
@@ -19,6 +19,7 @@ Depending on the architecture, different models handle reference images differen
 | **Krea2 (Community Edit LoRAs)** | `krea2_ostris_edit` |
 | [**Mage-Flow-Edit**](./mage_flow.md#image-editing) | `mage_flow` |
 | [**LLaDA-Image**](./llada_image.md#image-editing) | `llada_image` |
+| [**Fibo 2.0 Dev / Edit Turbo**](./fibo2.md#examples) | `fibo2` |
 | **Anima (Community Edit LoRAs)** | `cosmos_reference` |
 
 Stable-diffusion.spp also supports basic Unet-based editing models like instruct-pix2pix or CosXL-Edit. This document is not about those.
@@ -50,6 +51,7 @@ The `--ref-image-args` argument accepts a comma-separated list of key-value pair
 | `flux_kontext` | FLUX.1 Kontext |
 | `longcat` | LongCat Image Edit |
 | `flux2` | FLUX.2 models |
+| `fibo2` | Fibo 2.0 Dev / Edit Turbo |
 | `qwen` | Qwen Image Edit |
 | `qwen_layered` | Qwen Image Layered |
 | `z_image_omni` | Boogu, Z-Image Omni |
@@ -73,7 +75,12 @@ If presets are insufficient, you can manually configure the following parameters
 | `ref_index_mode` | string | Behavior of the RoPE index. | `fixed`, `increase`, `decrease` |
 | `force_ref_timestep_zero` | bool | Forces timestep=0 for reference tokens. | `true`, `false` (Krea2 only) |
 | `resize_before_vae` | bool | Whether reference images are resized before VAE encoding. | `true`, `false` |
-| `vae_input_max_pixels` | int | Maximum pixel area for VAE reference inputs. | Integer |
+| `resize_vae_to_target` | bool | Use output dimensions for every reference, overriding area-based sizing. | `true`, `false` |
+| `resize_single_ref_to_target` | bool | Use output dimensions when exactly one reference is supplied. | `true`, `false` |
+| `vae_size_buckets` | bool | Choose aligned dimensions within ±10% of the requested area, minimizing aspect-ratio cropping. Ignored when using output dimensions. | `true`, `false` |
+| `vae_center_crop` | bool | Scale to cover the selected dimensions with Lanczos, then center-crop. Otherwise resize directly. | `true`, `false` |
+| `vae_input_max_pixels` | int | VAE area budget, also capped by output area in ordinary resizing. With size buckets, this is the target area and output area does not cap it. | Integer |
+| `vlm_use_vae_input` | bool | Reuse the resized/cropped VAE input for the VLM, before the VLM's own resize. Has no effect when `resize_before_vae=false`. | `true`, `false` |
 | `vlm_resize_mode` | string | How to resize VLM reference inputs. | `longest_side`, `area`, `none` |
 | `vlm_max_size` | int | Maximum VLM input size; interpreted according to `vlm_resize_mode`. | Integer |
 | `vlm_min_size` | int | Minimum VLM input size; interpreted according to `vlm_resize_mode`. | Integer |
@@ -88,6 +95,7 @@ For a technical overview of how each preset is configured, see the table below.
 | `flux_kontext` | No | `fixed` | `none` | |
 | `longcat` | Yes | `fixed` | `area` | |
 | `flux2` | No | `increase` | `none` | |
+| `fibo2` | Yes | `increase` | `longest` | Shared reference crops; clean reference timesteps; see [Fibo 2.0](./fibo2.md) |
 | `qwen` | Yes | `increase` | `area` | |
 | `qwen_layered` | Yes | `decrease` | `area` | |
 | `mage_flow` | Yes | `increase` | `longest` | `vlm_max_size = 384`, VAE input resized to target |
@@ -99,3 +107,5 @@ For a technical overview of how each preset is configured, see the table below.
 **Additional Default Notes:**
 - **VLM Input Sizes:** For most presets, `vlm_max_size` and `vlm_min_size` are set to `-1`, meaning the values are model-dependent and handled automatically. In `area` mode they represent pixel area; in `longest_side` mode they represent a side length in pixels.
 - **VAE Input Size:** `vae_input_max_pixels` defaults to $1024 \times 1024$ pixels (`1048576`).
+- **VAE Geometry:** `resize_vae_to_target` defaults to `true` for Mage-Flow and LLaDA-Image. Fibo enables `resize_single_ref_to_target`, `vae_size_buckets`, and `vae_center_crop`. These options are otherwise `false`. Bucket dimensions use the model's alignment (normally 16, Qwen Image 32), with a minimum side of four alignment units; if no bucket fits, output dimensions are used.
+- **Shared Encoder Input:** `vlm_use_vae_input` defaults to `true` for Fibo and Qwen Image 2.1, otherwise `false`. All these options can be overridden individually after selecting a preset.
diff --git a/docs/fibo2.md b/docs/fibo2.md
@@ -0,0 +1,53 @@
+# How to Use
+
+Fibo 2.0 uses a Fibo diffusion transformer, the standard FLUX.2 VAE, and Qwen3-VL 4B as the LLM text encoder. The Qwen tokenizer is built in.
+
+## Download weights
+
+- Download Fibo 2.0 Dev
+    - safetensors: https://huggingface.co/Comfy-Org/Fibo-2.0/tree/main/diffusion_models
+- Download Fibo 2.0 Edit Turbo
+    - safetensors: https://huggingface.co/Comfy-Org/Fibo-2.0/tree/main/diffusion_models
+- Download vae
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-dev/tree/main
+- Download Qwen3-VL 4B
+    - safetensors: https://huggingface.co/Comfy-Org/Fibo-2.0/tree/main/text_encoders
+    - gguf: https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF/tree/main
+
+The original Diffusers weights from [briaai/Fibo-2.0-dev](https://huggingface.co/briaai/Fibo-2.0-dev) and [briaai/Fibo-2.0-edit-turbo](https://huggingface.co/briaai/Fibo-2.0-edit-turbo) are also supported. Pass the transformer's `diffusion_pytorch_model.safetensors.index.json` to `--diffusion-model` and keep its shards alongside it.
+
+INT8 convrot weights load directly; see [INT8 convrot](./int8_convrot.md) for backend requirements.
+
+## Examples
+
+### Fibo 2.0 Dev
+
+Dev supports both text-to-image generation and reference-image editing.
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\fibo2_dev_int8_convrot.safetensors --llm ..\models\text_encoders\Qwen3-VL-4B-Instruct-Q4_K_M.gguf --llm_vision ..\models\text_encoders\Qwen3-VL-4B-Instruct-mmproj-BF16.gguf --vae ..\models\vae\flux2_ae.safetensors -p "a lovely cat holding a sign says 'fibo2.cpp'" --cfg-scale 5.0 --sampling-method euler -v  --offload-to-cpu
+```
+
+<img src="../assets/fibo2/t2i.png" />
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\fibo2_dev_int8_convrot.safetensors --llm ..\models\text_encoders\Qwen3-VL-4B-Instruct-Q4_K_M.gguf --llm_vision ..\models\text_encoders\Qwen3-VL-4B-Instruct-mmproj-BF16.gguf --vae ..\models\vae\flux2_ae.safetensors -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'fibo2.cpp'" --cfg-scale 5.0 --sampling-method euler -v  --offload-to-cpu
+```
+
+<img src="../assets/fibo2/dev_edit.png" />
+
+### Fibo 2.0 Edit Turbo
+
+Edit Turbo is distilled for reference-image editing in 4 steps with CFG 1.
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\fibo2_edit_turbo_int8_convrot.safetensors --llm ..\models\text_encoders\Qwen3-VL-4B-Instruct-Q4_K_M.gguf --llm_vision ..\models\text_encoders\Qwen3-VL-4B-Instruct-mmproj-BF16.gguf --vae ..\models\vae\flux2_ae.safetensors -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'fibo2.cpp'" --cfg-scale 1 --steps 4 --sampling-method euler -v  --offload-to-cpu
+```
+
+The BF16 files accept the same options. Use dimensions divisible by 16. Repeat `-r` to use multiple reference images; instructions can refer to `<image_1>`, `<image_2>`, etc. See [Image Editing](./edit.md) for reference-image preprocessing options. Separate GGUF text encoders need their vision projector passed with `--llm_vision` for editing.
+
+The automatic `fibo2` reference preset combines `resize_single_ref_to_target=true`, `vae_size_buckets=true`, `vae_center_crop=true`, and `vlm_use_vae_input=true`. A single reference is scaled to cover the output dimensions and center-cropped. Multiple references use dimensions divisible by 16, chosen within ±10% of one megapixel to minimize cropping for each image's aspect ratio; `vae_input_max_pixels` overrides that target area. The VAE and vision encoder receive the same crops, with smaller copies used by the vision encoder. These are generic reference-image options and can be overridden individually: for example, `--ref-image-args vae_center_crop=false` uses direct resizing, while `resize_before_vae=false` skips this automatic crop and resize step entirely.
+
+The vision encoder defaults to `vlm_resize_mode=longest_side,vlm_max_size=768`. Override these with `--ref-image-args`, for example `vlm_max_size=512`, or use `vlm_resize_mode=none` to skip size-based scaling. Vision inputs are still aligned to the encoder's patch grid.
+
+Plain prompts and structured JSON captions are accepted. This integration uses the caption-only text-rendering path (`render_text_images=False`); it does not generate auxiliary glyph images or expand prompts with Bria VGL. Native Fibo mask preprocessing is not connected to `--mask`.
diff --git a/docs/performance.md b/docs/performance.md
@@ -32,8 +32,10 @@ segment count change.
 path. Performance depends on GPU architecture, backend, driver, model, and
 tensor shapes, so enabling direct convolution can either improve or reduce speed.
 
-Automatic selection uses small benchmarks on Vulkan and implicit GEMM checks
-on other backends. It does not guarantee the fastest end-to-end generation.
+Automatic selection uses small benchmarks on Vulkan, implicit GEMM checks on
+CUDA/HIP/MUSA, and the direct kernels on CPU when ggml carries the row-run
+patch (see [backend selection](backend.md)). It does not guarantee the fastest
+end-to-end generation.
 Compare `on` and `off` with the same model and generation settings over
 repeated runs, keeping flash attention unchanged and separating startup/probe
 time from steady-state generation time. See [backend selection](backend.md)