FD-Loss-diffusers

Post-training checkpoints with the same JiT / iMF / pMF architectures as the base models, distilled with FD-loss (feature distillation).

• JiT: all /16 variants for 256px generation
• iMF: all variants for 256px generation
• pMF: both /16 and /32
• SIM suffix: SigLIP + Inception + MAE FD-loss
• Default inference: 1 NFE (num_inference_steps=1)

FD-Loss JiT uses legacy time convention and velocity Euler via bundled FDLossFlowMatchScheduler (scheduler/scheduling_flow_match_fd.py; timesteps t=1→0).

Demo

Class-conditional sample (ImageNet class golden retriever), JiT-H/16 FD-SIM at 256×256, 1 NFE, CFG 2.2, interval [0.1, 1.0], seed 42.

Inference (JiT-B-16-SIM)

from pathlib import Path
import torch
from diffusers import DiffusionPipeline

model_dir = Path("./JiT-B-16-SIM")
pipe = DiffusionPipeline.from_pretrained(
    str(model_dir),
    custom_pipeline=str(model_dir / "pipeline.py"),
    trust_remote_code=True,
)
pipe.to("cuda")

generator = torch.Generator(device="cuda").manual_seed(42)
image = pipe(
    class_labels="golden retriever",
    num_inference_steps=1,   # 1 NFE (FD-Loss default)
    guidance_scale=3.0,
    guidance_interval_min=0.1,
    guidance_interval_max=1.0,
    generator=generator,
).images[0]

Parameter	JiT-B-16-SIM default	Source
num_inference_steps	1	--num_sampling_steps 1
guidance_scale	3.0	JiT_B eval preset
guidance_interval_min / max	0.1 / 1.0	JiT_B eval preset
legacy_time_convention	True (pipeline default)	--legacy_time_convention

Inference (JiT-L-16-SIM)

Same loading pattern as JiT-B; use CFG 2.4 (JiT_L eval preset):

model_dir = Path("./JiT-L-16-SIM")
pipe = DiffusionPipeline.from_pretrained(
    str(model_dir),
    custom_pipeline=str(model_dir / "pipeline.py"),
    trust_remote_code=True,
).to("cuda")

image = pipe(
    class_labels="golden retriever",
    num_inference_steps=1,
    guidance_scale=2.4,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]

Inference (JiT-H-16-SIM)

Same loading pattern as JiT-B/L; use CFG 2.2 (JiT_H eval preset):

model_dir = Path("./JiT-H-16-SIM")
pipe = DiffusionPipeline.from_pretrained(
    str(model_dir),
    custom_pipeline=str(model_dir / "pipeline.py"),
    trust_remote_code=True,
).to("cuda")

image = pipe(
    class_labels="golden retriever",
    num_inference_steps=1,
    guidance_scale=2.2,
    guidance_interval_min=0.1,
    guidance_interval_max=1.0,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]

Available variants

Variant	Path	Architecture	Resolution	CFG (1 NFE)
JiT-B-16-SIM	./JiT-B-16-SIM	JiT-B/16	256×256	3.0
JiT-L-16-SIM	./JiT-L-16-SIM	JiT-L/16	256×256	2.4
JiT-H-16-SIM	./JiT-H-16-SIM	JiT-H/16	256×256	2.2
iMF-B-SIM	./iMF-B-SIM	iMF-B/2	256×256	8.0
iMF-L-SIM	./iMF-L-SIM	iMF-L/2	256×256	8.0
iMF-XL-SIM	./iMF-XL-SIM	iMF-XL/2	256×256	8.0
pMF-B-16-SIM	./pMF-B-16-SIM	pMF-B/16	256×256	7.5
pMF-B-32-SIM	./pMF-B-32-SIM	pMF-B/32	512×512	6.5
pMF-L-16-SIM	./pMF-L-16-SIM	pMF-L/16	256×256	7.0
pMF-L-32-SIM	./pMF-L-32-SIM	pMF-L/32	512×512	7.5
pMF-H-32-SIM	./pMF-H-32-SIM	pMF-H/32	512×512	5.5

Inference (iMF-B-SIM)

Uses production IMFPipeline from iMF-diffusers/iMF-B-2 (native iMF time convention, not legacy JiT time). Use torch.float32 (same as base iMF variants):

model_dir = Path("./iMF-B-SIM")
pipe = DiffusionPipeline.from_pretrained(
    str(model_dir),
    custom_pipeline=str(model_dir / "pipeline.py"),
    trust_remote_code=True,
    torch_dtype=torch.float32,
).to("cuda")

image = pipe(
    class_labels="golden retriever",
    num_inference_steps=1,
    guidance_scale=8.0,
    guidance_interval_start=0.4,
    guidance_interval_end=0.65,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]

Inference (iMF-L-SIM)

Same loading pattern as iMF-B-SIM (production IMFPipeline from iMF-diffusers/iMF-L-2):

model_dir = Path("./iMF-L-SIM")
pipe = DiffusionPipeline.from_pretrained(
    str(model_dir),
    custom_pipeline=str(model_dir / "pipeline.py"),
    trust_remote_code=True,
    torch_dtype=torch.float32,
).to("cuda")

image = pipe(
    class_labels="golden retriever",
    num_inference_steps=1,
    guidance_scale=8.0,
    guidance_interval_start=0.4,
    guidance_interval_end=0.65,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]

Regenerate from checkpoint:

python _convert_imf_l_fd_sim.py

Inference (iMF-XL-SIM)

Same loading pattern as iMF-B/L-SIM (production IMFPipeline from iMF-diffusers/iMF-XL-2):

model_dir = Path("./iMF-XL-SIM")
pipe = DiffusionPipeline.from_pretrained(
    str(model_dir),
    custom_pipeline=str(model_dir / "pipeline.py"),
    trust_remote_code=True,
    torch_dtype=torch.float32,
).to("cuda")

image = pipe(
    class_labels="golden retriever",
    num_inference_steps=1,
    guidance_scale=8.0,
    guidance_interval_start=0.4,
    guidance_interval_end=0.65,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]

Inference (pMF-B-16-SIM)

Uses production PMFPipeline from pMF-diffusers/pMF-B-16 (native pMF time convention). Use torch.bfloat16 on CUDA:

model_dir = Path("./pMF-B-16-SIM")
pipe = DiffusionPipeline.from_pretrained(
    str(model_dir),
    custom_pipeline=str(model_dir / "pipeline.py"),
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
).to("cuda")

image = pipe(
    class_labels="golden retriever",
    num_inference_steps=1,
    guidance_scale=7.5,
    guidance_interval_min=0.1,
    guidance_interval_max=0.8,
    noise_scale=1.0,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]

Inference (pMF-B-32-SIM)

Same loading pattern as pMF-B-16-SIM (production PMFPipeline from pMF-diffusers/pMF-B-32):

model_dir = Path("./pMF-B-32-SIM")
pipe = DiffusionPipeline.from_pretrained(
    str(model_dir),
    custom_pipeline=str(model_dir / "pipeline.py"),
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
).to("cuda")

image = pipe(
    class_labels="golden retriever",
    num_inference_steps=1,
    guidance_scale=6.5,
    guidance_interval_min=0.1,
    guidance_interval_max=0.7,
    noise_scale=2.0,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]