Skip to content

Feat; Add support for Wan/Qwen TAEHV decoding #937

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
wants to merge 5 commits into
base: master
Choose a base branch
from
+373 −50
Draft

Feat; Add support for Wan/Qwen TAEHV decoding #937

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
fe1f6e0
Add support for Wan2.1 TAEHV decoding
stduhpf Nov 3, 2025
60bec38
main: --tae instead of --taesd
stduhpf Nov 3, 2025
ea34cdb
progress towards video support
stduhpf Nov 3, 2025
6795acc
Wan2.1 decode not crashing anymore (still broken)
stduhpf Nov 3, 2025
d2968b1
Less broken video decode + remove log spam
stduhpf Nov 4, 2025
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 examples/cli/main.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -67,7 +67,7 @@ struct SDParams {
std::string diffusion_model_path;
std::string high_noise_diffusion_model_path;
std::string vae_path;
std::string taesd_path;
std::string tae_path;
std::string esrgan_path;
std::string control_net_path;
std::string embedding_dir;
Expand Down Expand Up @@ -159,7 +159,7 @@ void print_params(SDParams params) {
printf(" diffusion_model_path: %s\n", params.diffusion_model_path.c_str());
printf(" high_noise_diffusion_model_path: %s\n", params.high_noise_diffusion_model_path.c_str());
printf(" vae_path: %s\n", params.vae_path.c_str());
printf(" taesd_path: %s\n", params.taesd_path.c_str());
printf(" tae_path: %s\n", params.tae_path.c_str());
printf(" esrgan_path: %s\n", params.esrgan_path.c_str());
printf(" control_net_path: %s\n", params.control_net_path.c_str());
printf(" embedding_dir: %s\n", params.embedding_dir.c_str());
Expand Down Expand Up @@ -523,10 +523,10 @@ void parse_args(int argc, const char** argv, SDParams& params) {
"--vae",
"path to standalone vae model",
&params.vae_path},
{"",
{"--tae",
"--taesd",
"path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)",
&params.taesd_path},
"path to taesd or taehv. Using Tiny AutoEncoder for fast decoding (low quality)",
&params.tae_path},
{"",
"--control-net",
"path to control net model",
Expand Down Expand Up @@ -1638,7 +1638,7 @@ int main(int argc, const char* argv[]) {
params.diffusion_model_path.c_str(),
params.high_noise_diffusion_model_path.c_str(),
params.vae_path.c_str(),
params.taesd_path.c_str(),
params.tae_path.c_str(),
params.control_net_path.c_str(),
params.lora_model_dir.c_str(),
params.embedding_dir.c_str(),
Expand Down
95 changes: 58 additions & 37 deletions stable-diffusion.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -346,8 +346,8 @@ class StableDiffusionGGML {
offload_params_to_cpu,
tensor_storage_map);
diffusion_model = std::make_shared<MMDiTModel>(backend,
offload_params_to_cpu,
tensor_storage_map);
offload_params_to_cpu,
tensor_storage_map);
} else if (sd_version_is_flux(version)) {
bool is_chroma = false;
for (auto pair : tensor_storage_map) {
Expand Down Expand Up @@ -389,10 +389,10 @@ class StableDiffusionGGML {
1,
true);
diffusion_model = std::make_shared<WanModel>(backend,
offload_params_to_cpu,
tensor_storage_map,
"model.diffusion_model",
version);
offload_params_to_cpu,
tensor_storage_map,
"model.diffusion_model",
version);
if (strlen(SAFE_STR(sd_ctx_params->high_noise_diffusion_model_path)) > 0) {
high_noise_diffusion_model = std::make_shared<WanModel>(backend,
offload_params_to_cpu,
Expand All @@ -418,10 +418,10 @@ class StableDiffusionGGML {
"",
enable_vision);
diffusion_model = std::make_shared<QwenImageModel>(backend,
offload_params_to_cpu,
tensor_storage_map,
"model.diffusion_model",
version);
offload_params_to_cpu,
tensor_storage_map,
"model.diffusion_model",
version);
} else { // SD1.x SD2.x SDXL
if (strstr(SAFE_STR(sd_ctx_params->photo_maker_path), "v2")) {
cond_stage_model = std::make_shared<FrozenCLIPEmbedderWithCustomWords>(clip_backend,
Expand Down Expand Up @@ -475,14 +475,27 @@ class StableDiffusionGGML {
}

if (sd_version_is_wan(version) || sd_version_is_qwen_image(version)) {
first_stage_model = std::make_shared<WAN::WanVAERunner>(vae_backend,
offload_params_to_cpu,
tensor_storage_map,
"first_stage_model",
vae_decode_only,
version);
first_stage_model->alloc_params_buffer();
first_stage_model->get_param_tensors(tensors, "first_stage_model");
if (!use_tiny_autoencoder) {
first_stage_model = std::make_shared<WAN::WanVAERunner>(vae_backend,
offload_params_to_cpu,
tensor_storage_map,
"first_stage_model",
vae_decode_only,
version);
first_stage_model->alloc_params_buffer();
first_stage_model->get_param_tensors(tensors, "first_stage_model");
} else {
tae_first_stage = std::make_shared<TinyVideoAutoEncoder>(vae_backend,
offload_params_to_cpu,
tensor_storage_map,
"decoder",
vae_decode_only,
version);
if (sd_ctx_params->vae_conv_direct) {
LOG_INFO("Using Conv2d direct in the tae model");
tae_first_stage->set_conv2d_direct_enabled(true);
}
}
} else if (version == VERSION_CHROMA_RADIANCE) {
first_stage_model = std::make_shared<FakeVAE>(vae_backend,
offload_params_to_cpu);
Expand Down Expand Up @@ -510,12 +523,12 @@ class StableDiffusionGGML {
first_stage_model->alloc_params_buffer();
first_stage_model->get_param_tensors(tensors, "first_stage_model");
} else {
tae_first_stage = std::make_shared<TinyAutoEncoder>(vae_backend,
offload_params_to_cpu,
tensor_storage_map,
"decoder.layers",
vae_decode_only,
version);
tae_first_stage = std::make_shared<TinyImageAutoEncoder>(vae_backend,
offload_params_to_cpu,
tensor_storage_map,
"decoder.layers",
vae_decode_only,
version);
if (sd_ctx_params->vae_conv_direct) {
LOG_INFO("Using Conv2d direct in the tae model");
tae_first_stage->set_conv2d_direct_enabled(true);
Expand Down Expand Up @@ -625,12 +638,15 @@ class StableDiffusionGGML {
unet_params_mem_size += high_noise_diffusion_model->get_params_buffer_size();
}
size_t vae_params_mem_size = 0;
LOG_DEBUG("Here");
if (!use_tiny_autoencoder) {
vae_params_mem_size = first_stage_model->get_params_buffer_size();
} else {
LOG_DEBUG("Here");
if (!tae_first_stage->load_from_file(taesd_path, n_threads)) {
return false;
}
LOG_DEBUG("Here");
vae_params_mem_size = tae_first_stage->get_params_buffer_size();
}
size_t control_net_params_mem_size = 0;
Expand Down Expand Up @@ -1428,12 +1444,12 @@ class StableDiffusionGGML {
-0.0313f, -0.1649f, 0.0117f, 0.0723f, -0.2839f, -0.2083f, -0.0520f, 0.3748f,
0.0152f, 0.1957f, 0.1433f, -0.2944f, 0.3573f, -0.0548f, -0.1681f, -0.0667f};
latents_std_vec = {
0.4765f, 1.0364f, 0.4514f, 1.1677f, 0.5313f, 0.4990f, 0.4818f, 0.5013f,
0.8158f, 1.0344f, 0.5894f, 1.0901f, 0.6885f, 0.6165f, 0.8454f, 0.4978f,
0.5759f, 0.3523f, 0.7135f, 0.6804f, 0.5833f, 1.4146f, 0.8986f, 0.5659f,
0.7069f, 0.5338f, 0.4889f, 0.4917f, 0.4069f, 0.4999f, 0.6866f, 0.4093f,
0.5709f, 0.6065f, 0.6415f, 0.4944f, 0.5726f, 1.2042f, 0.5458f, 1.6887f,
0.3971f, 1.0600f, 0.3943f, 0.5537f, 0.5444f, 0.4089f, 0.7468f, 0.7744f};
0.4765f, 1.0364f, 0.4514f, 1.1677f, 0.5313f, 0.4990f, 0.4818f, 0.5013f,
0.8158f, 1.0344f, 0.5894f, 1.0901f, 0.6885f, 0.6165f, 0.8454f, 0.4978f,
0.5759f, 0.3523f, 0.7135f, 0.6804f, 0.5833f, 1.4146f, 0.8986f, 0.5659f,
0.7069f, 0.5338f, 0.4889f, 0.4917f, 0.4069f, 0.4999f, 0.6866f, 0.4093f,
0.5709f, 0.6065f, 0.6415f, 0.4944f, 0.5726f, 1.2042f, 0.5458f, 1.6887f,
0.3971f, 1.0600f, 0.3943f, 0.5537f, 0.5444f, 0.4089f, 0.7468f, 0.7744f};
}
for (int i = 0; i < latent->ne[3]; i++) {
float mean = latents_mean_vec[i];
Expand Down Expand Up @@ -1474,12 +1490,12 @@ class StableDiffusionGGML {
-0.0313f, -0.1649f, 0.0117f, 0.0723f, -0.2839f, -0.2083f, -0.0520f, 0.3748f,
0.0152f, 0.1957f, 0.1433f, -0.2944f, 0.3573f, -0.0548f, -0.1681f, -0.0667f};
latents_std_vec = {
0.4765f, 1.0364f, 0.4514f, 1.1677f, 0.5313f, 0.4990f, 0.4818f, 0.5013f,
0.8158f, 1.0344f, 0.5894f, 1.0901f, 0.6885f, 0.6165f, 0.8454f, 0.4978f,
0.5759f, 0.3523f, 0.7135f, 0.6804f, 0.5833f, 1.4146f, 0.8986f, 0.5659f,
0.7069f, 0.5338f, 0.4889f, 0.4917f, 0.4069f, 0.4999f, 0.6866f, 0.4093f,
0.5709f, 0.6065f, 0.6415f, 0.4944f, 0.5726f, 1.2042f, 0.5458f, 1.6887f,
0.3971f, 1.0600f, 0.3943f, 0.5537f, 0.5444f, 0.4089f, 0.7468f, 0.7744f};
0.4765f, 1.0364f, 0.4514f, 1.1677f, 0.5313f, 0.4990f, 0.4818f, 0.5013f,
0.8158f, 1.0344f, 0.5894f, 1.0901f, 0.6885f, 0.6165f, 0.8454f, 0.4978f,
0.5759f, 0.3523f, 0.7135f, 0.6804f, 0.5833f, 1.4146f, 0.8986f, 0.5659f,
0.7069f, 0.5338f, 0.4889f, 0.4917f, 0.4069f, 0.4999f, 0.6866f, 0.4093f,
0.5709f, 0.6065f, 0.6415f, 0.4944f, 0.5726f, 1.2042f, 0.5458f, 1.6887f,
0.3971f, 1.0600f, 0.3943f, 0.5537f, 0.5444f, 0.4089f, 0.7468f, 0.7744f};
}
for (int i = 0; i < latent->ne[3]; i++) {
float mean = latents_mean_vec[i];
Expand Down Expand Up @@ -1706,6 +1722,10 @@ class StableDiffusionGGML {
first_stage_model->free_compute_buffer();
process_vae_output_tensor(result);
} else {
if (sd_version_is_wan(version)) {
x = ggml_permute(work_ctx, x, 0, 1, 3, 2);
}

if (vae_tiling_params.enabled && !decode_video) {
// split latent in 64x64 tiles and compute in several steps
auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
Expand All @@ -1716,6 +1736,7 @@ class StableDiffusionGGML {
tae_first_stage->compute(n_threads, x, true, &result);
}
tae_first_stage->free_compute_buffer();

}

int64_t t1 = ggml_time_ms();
Expand Down Expand Up @@ -3104,7 +3125,7 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
struct ggml_tensor* vid = sd_ctx->sd->decode_first_stage(work_ctx, final_latent, true);
int64_t t5 = ggml_time_ms();
LOG_INFO("decode_first_stage completed, taking %.2fs", (t5 - t4) * 1.0f / 1000);
if (sd_ctx->sd->free_params_immediately) {
if (sd_ctx->sd->free_params_immediately && !sd_ctx->sd->use_tiny_autoencoder) {
sd_ctx->sd->first_stage_model->free_params_buffer();
}

Expand Down
Loading
Loading