Skip to content

Adds Mage Flow support #1454

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
jtreminio wants to merge 1 commit into
base: master
Choose a base branch
from
Draft

Adds Mage Flow support #1454

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 6 additions & 6 deletions src/BuiltinExtensions/ComfyUIBackend/WorkflowGenerator.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -769,9 +769,9 @@ public List<JArray> LoadPromptImagesForMainRef(List<Image> images)
public (JArray, JArray, JArray, JArray) BuildInputImageHandling(List<JArray> images, JArray pos, JArray neg, JArray latent)
{
JArray imgNeg = null;
if (IsKontext() || IsOmniGen() || IsQwenImage() || IsAnyFlux2() || IsBoogu())
if (IsKontext() || IsOmniGen() || IsQwenImage() || IsAnyFlux2() || IsBoogu() || IsMageFlow())
{
if (IsOmniGen() || IsQwenImageEditPlus() || IsBoogu())
if (IsOmniGen() || IsQwenImageEditPlus() || IsBoogu() || IsMageFlow())
{
imgNeg = neg;
}
Expand Down Expand Up @@ -803,7 +803,7 @@ void makeRefLatent(JArray image)
}
if (img is not null)
{
if (IsQwenImageEditPlus() || IsBoogu())
if (IsQwenImageEditPlus() || IsBoogu() || IsMageFlow())
{
neg = imgNeg;
}
Expand Down Expand Up @@ -975,7 +975,7 @@ public string CreateKSampler(JArray model, JArray pos, JArray neg, JArray latent
defscheduler ??= "simple";
}
// TODO: Registry of model default preferences instead of this
else if (IsFlux() || IsWanVideo() || IsWanVideo22() || IsOmniGen() || IsQwenImage() || IsZImage() || IsZetaChroma() || IsErnie() || IsHiDreamO1() || IsLens() || IsPixelDiT() || IsKrea2())
else if (IsFlux() || IsWanVideo() || IsWanVideo22() || IsOmniGen() || IsQwenImage() || IsZImage() || IsZetaChroma() || IsErnie() || IsHiDreamO1() || IsLens() || IsPixelDiT() || IsKrea2() || IsMageFlow())
{
defscheduler ??= "simple";
}
Expand Down Expand Up @@ -1305,7 +1305,7 @@ public JArray GetPromptImage(bool fixSize, bool promptSize = false, int index =
doesFit = false;
}
}
else if ((IsBoogu() || IsQwenImageEditPlus()) && promptSize)
else if ((IsBoogu() || IsQwenImageEditPlus() || IsMageFlow()) && promptSize)
{
target = 384;
doesFit = false;
Expand Down Expand Up @@ -2353,7 +2353,7 @@ public JArray CreateConditioningDirect(string prompt, JArray clip, T2IModel mode
["text"] = prompt
}, id);
}
else if (IsIdeogram4() || IsKrea2() || (IsBoogu() && isPositive))
else if (IsIdeogram4() || IsKrea2() || (IsBoogu() && isPositive) || IsMageFlow())
{
JArray imageNode = GetPromptImage(true, true, 0);
for (int i = 1; i < 10; i++)
Expand Down
8 changes: 8 additions & 0 deletions src/BuiltinExtensions/ComfyUIBackend/WorkflowGeneratorModelSupport.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -218,6 +218,9 @@ public bool IsNvidiaCosmos2()
/// <summary>Returns true if the current model is any Kandinsky 5 variant.</summary>
public bool IsKandinsky5() => IsKandinsky5ImgLite() || IsKandinsky5VidLite() || IsKandinsky5VidPro();

/// <summary>Returns true if the current model is Mage-Flow.</summary>
public bool IsMageFlow() => IsModelCompatClass(T2IModelClassSorter.CompatMageFlow);

/// <summary>Returns true if the current model is any Wan-2.1 variant.</summary>
public bool IsWanVideo()
{
Expand Down Expand Up @@ -1169,6 +1172,11 @@ public void LoadClip3(string type, string modelA, string modelB, string modelC)
helpers.LoadClip("boogu", helpers.GetQwen3vl_8bModel());
helpers.DoVaeLoader(UserInput.SourceSession?.User?.Settings?.VAEs?.DefaultFluxVAE, "flux-1", "flux-ae");
}
else if (IsMageFlow())
{
helpers.LoadClip("mage", helpers.GetQwen3vl_4bModel());
helpers.DoVaeLoader(null, "mage-flow", "mage-flow-vae");
}
else if (IsFlux() && (LoadingClip is null || LoadingVAE is null || UserInput.Get(T2IParamTypes.T5XXLModel) is not null || UserInput.Get(T2IParamTypes.ClipLModel) is not null))
{
helpers.LoadClip2("flux", helpers.GetT5XXLModel(), helpers.GetClipLModel());
Expand Down
1 change: 1 addition & 0 deletions src/Text2Image/CommonModels.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,6 +92,7 @@ public static void RegisterCoreSet()
Register(new("ltx2-3-video-vae", "LTX-2.3 Video VAE", "The video VAE for Lightricks LTX-2.3.", "https://huggingface.co/Kijai/LTX2.3_comfy/resolve/main/vae/LTX23_video_vae_bf16.safetensors", "01ea62d09bc139f95c5dee7b5c062ad6a3e6cd8be910a1983ac02e7eb5b8ee3b", "VAE", "LTX-2/LTX23_video_vae_bf16.safetensors"));
Register(new("ltx2-3-audio-vae", "LTX-2.3 Audio VAE", "The audio VAE for Lightricks LTX-2.3.", "https://huggingface.co/Kijai/LTX2.3_comfy/resolve/main/vae/LTX23_audio_vae_bf16.safetensors", "5bc10fa4adecf99dda132d916e23048cbd56797702c5fa50eb5d2079048a38c3", "VAE", "LTX-2/LTX23_audio_vae_bf16.safetensors"));
Register(new("qwen-image-vae", "Qwen Image VAE", "The VAE for Qwen Image", "https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/vae/qwen_image_vae.safetensors", "a70580f0213e67967ee9c95f05bb400e8fb08307e017a924bf3441223e023d1f", "VAE", "QwenImage/qwen_image_vae.safetensors"));
Register(new("mage-flow-vae", "Mage-VAE", "The VAE for Mage-Flow", "https://huggingface.co/microsoft/Mage-Flow/resolve/main/vae/diffusion_pytorch_model.safetensors", "34e076dc1e8a15321e1e07be5111d59cf16dd10b804b7c7e20b4de29013427e0", "VAE", "MageFlow/mage_flow_vae.safetensors"));
Register(new("hunyuan-image-2_1-vae", "Hunyuan Image 2.1 VAE", "The VAE for Hunyuan Image 2.1 Base", "https://huggingface.co/Comfy-Org/HunyuanImage_2.1_ComfyUI/resolve/main/split_files/vae/hunyuan_image_2.1_vae_fp16.safetensors", "f2ae19863609206196b5e3a86bfd94f67bd3866f5042004e3994f07e3c93b2f9", "VAE", "HunyuanImage/hunyuan_image_2.1_vae_fp16.safetensors"));
Register(new("hunyuan-image-2_1-refiner-vae", "Hunyuan Image 2.1 Refiner VAE", "The VAE for Hunyuan Image 2.1 Refiner", "https://huggingface.co/Comfy-Org/HunyuanImage_2.1_ComfyUI/resolve/main/split_files/vae/hunyuan_image_refiner_vae_fp16.safetensors", "e1b74e85d61b65e18cc05ca390e387d93cfadf161e737de229ebb800ea3db769", "VAE", "HunyuanImage/hunyuan_image_2.1_refiner_vae_fp16.safetensors"));
Register(new("hunyuan-video-1_5-vae", "Hunyuan Video 1.5 VAE", "The VAE for Hunyuan Video 1.5", "https://huggingface.co/Comfy-Org/HunyuanVideo_1.5_repackaged/resolve/main/split_files/vae/hunyuanvideo15_vae_fp16.safetensors", "e7c3091949c27e2d55ae6d5df917b99dadfebbf308e5a50d0ade0d16c90297ae", "VAE", "HunyuanVideo/hunyuanvideo15_vae_fp16.safetensors"));
Expand Down
9 changes: 8 additions & 1 deletion src/Text2Image/T2IModelClassSorter.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -95,6 +95,7 @@ public static T2IModelCompatClass
CompatIdeogram4 = RegisterCompat(new() { ID = "ideogram-4", ShortCode = "Ideo4", LorasTargetTextEnc = false, VaeFamily = VaeFlux2 }),
CompatKrea2 = RegisterCompat(new() { ID = "krea-2", ShortCode = "Krea2", VaeFamily = VaeQwenImage }),
CompatBoogu = RegisterCompat(new() { ID = "boogu", ShortCode = "Boogu", LorasTargetTextEnc = false, VaeFamily = VaeFlux1 }),
CompatMageFlow = RegisterCompat(new() { ID = "mage-flow", ShortCode = "Mage", LorasTargetTextEnc = false, VaeFamily = VaeFlux2 }),
// Audio models
CompatAceStep15 = RegisterCompat(new() { ID = "ace-step-1_5", ShortCode = "Ace15", IsAudioModel = true }),
// Obscure old random ones
Expand Down Expand Up @@ -245,6 +246,7 @@ bool isZImageLora(JObject h) => (hasLoraKey(h, "layers.0.adaLN_modulation.0") &&
bool isQwenImage(JObject h) => (h.ContainsKey("time_text_embed.timestep_embedder.linear_1.bias") && h.ContainsKey("img_in.bias") && (h.ContainsKey("transformer_blocks.0.attn.add_k_proj.bias") || h.ContainsKey("transformer_blocks.0.attn.add_qkv_proj.bias")))
|| (h.ContainsKey("model.diffusion_model.time_text_embed.timestep_embedder.linear_1.bias") && h.ContainsKey("model.diffusion_model.img_in.bias") && (h.ContainsKey("model.diffusion_model.transformer_blocks.0.attn.add_k_proj.bias") || h.ContainsKey("model.diffusion_model.transformer_blocks.0.attn.add_qkv_proj.bias")));
bool isQwenImageEdit2511(JObject h) => h.ContainsKey("__index_timestep_zero__");
bool isMageFlow(JObject h) => tryGetKey(h, "txt_norm.weight", out JToken tok) && tok["shape"].ToArray()[0].Value<long>() == 2560;

@jtreminio jtreminio Jul 22, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Qwen's txt_norm.weight is 3584.

Let me know if you need more specificity

bool isQwenImageLora(JObject h) => (hasLoraKey(h, "transformer_blocks.0.attn.add_k_proj") && hasLoraKey(h, "transformer_blocks.0.img_mlp.net.0.proj"))
|| (hasLoraKey(h, "transformer_blocks.0.attn.add_k_proj") && hasLoraKey(h, "transformer_blocks.0.img_mlp.net.2"))
|| (hasLoraKey(h, "transformer_blocks.59.attn.to_k") && hasLoraKey(h, "transformer_blocks.59.attn.to_out.0"))
Expand Down Expand Up @@ -666,7 +668,7 @@ JToken GetEmbeddingKey(JObject h)
// ====================== Qwen Image ======================
Register(new() { ID = "qwen-image", CompatClass = CompatQwenImage, Name = "Qwen Image", StandardWidth = 1328, StandardHeight = 1328, IsThisModelOfClass = (m, h) =>
{
return isQwenImage(h) && !isControlnetX(h) && !isSD3Controlnet(h) && !isQwenImageEdit2511(h);
return isQwenImage(h) && !isControlnetX(h) && !isSD3Controlnet(h) && !isQwenImageEdit2511(h) && !isMageFlow(h);
}});
Register(new() { ID = "qwen-image-edit", CompatClass = CompatQwenImage, Name = "Qwen Image Edit", StandardWidth = 1328, StandardHeight = 1328, IsThisModelOfClass = (m, h) =>
{
Expand Down Expand Up @@ -697,6 +699,11 @@ JToken GetEmbeddingKey(JObject h)
{
return isErnie(h);
}});
// ====================== Mage Flow ======================
Register(new() { ID = "mage-flow", CompatClass = CompatMageFlow, Name = "Mage Flow", StandardWidth = 1024, StandardHeight = 1024, IsThisModelOfClass = (m, h) =>
{
return isMageFlow(h) && isQwenImage(h) && !isControlnetX(h) && !isSD3Controlnet(h) && !isQwenImageEdit2511(h);
}});
// ====================== Kandinsky5 ======================
Register(new() { ID = "kandinsky5-image-lite", CompatClass = CompatKandinsky5ImgLite, Name = "Kandinsky5 Image Lite", StandardWidth = 1024, StandardHeight = 1024, IsThisModelOfClass = (m, h) =>
{
Expand Down
Loading