Comfy-Org/ComfyUI · error · ValueError

Unsupported RIFE model: expected 5 blocks, found {len(channe

Error message

Unsupported RIFE model: expected 5 blocks, found {len(channels)}

What it means

Raised by detect_rife_config() in comfy_extras/frame_interpolation_models/ifnet.py when probing a RIFE checkpoint's state dict: it reads the encoder channel count and then scans blocks.0..4.conv0.1.0.weight, requiring exactly 5 residual blocks with the expected key layout. Fewer than 5 matching keys means the checkpoint is a different RIFE architecture generation (RIFE v2/v4/v4.x variants have different block counts or key names) and this IFNet implementation cannot load it.

Source

Thrown at comfy_extras/frame_interpolation_models/ifnet.py:130

            else:
                fd, mask, feat = block(
                    torch.cat((warped_img0, warped_img1, self.warp(f0, flow[:, :2]), self.warp(f1, flow[:, 2:4]), timestep, mask, feat), 1),
                    flow, scale=self.scale_list[i])
                flow = flow.add_(fd)
            warped_img0 = self.warp(img0, flow[:, :2])
            warped_img1 = self.warp(img1, flow[:, 2:4])
        return torch.lerp(warped_img1, warped_img0, torch.sigmoid(mask))


def detect_rife_config(state_dict):
    head_ch = state_dict["encode.cnn3.weight"].shape[1]  # ConvTranspose2d: (in_ch, out_ch, kH, kW)
    channels = []
    for i in range(5):
        key = f"blocks.{i}.conv0.1.0.weight"
        if key in state_dict:
            channels.append(state_dict[key].shape[0])
    if len(channels) != 5:
        raise ValueError(f"Unsupported RIFE model: expected 5 blocks, found {len(channels)}")
    return head_ch, channels

View on GitHub (pinned to 1c6d8d45b3)

Solutions

  1. Use a RIFE v4-family checkpoint known to work with ComfyUI's frame interpolation node (the officially referenced 4.x weights).
  2. Inspect the state dict keys (torch.load then list(sd.keys())) and confirm blocks.0-4.conv0.1.0.weight all exist; if prefixed with 'module.', strip the prefix before load.
  3. If the checkpoint is a different architecture version, convert it or obtain the matching version's weights rather than expecting this loader to handle it.
  4. Verify integrity of the download — truncated files can load partially with missing keys.

Example fix

# before
sd = torch.load('rife_v2.pth')
detect_rife_config(sd)  # ValueError: expected 5 blocks, found 0

# after: use supported v4-family weights, and normalize prefixes if needed
sd = torch.load('rife46.pth')
sd = { k[len('module.'):] if k.startswith('module.') else k: v for k, v in sd.items() }
detect_rife_config(sd)
Defensive patterns

Strategy: validation

Validate before calling

import torch

def is_supported_rife(sd) -> bool:
    if 'encode.cnn3.weight' not in sd: return False
    keys = [f'blocks.{i}.conv0.1.0.weight' for i in range(5)]
    return all(k in sd for k in keys)

sd = torch.load(path, map_location='cpu')
assert is_supported_rife(sd), 'checkpoint is not a supported RIFE v4-family model'

Type guard

def is_supported_rife(sd) -> bool:
    if 'encode.cnn3.weight' not in sd: return False
    return all(f'blocks.{i}.conv0.1.0.weight' in sd for i in range(5))

Try / catch

try:
    head_ch, channels = detect_rife_config(sd)
except ValueError as e:
    raise RuntimeError(f'Incompatible RIFE checkpoint ({path}): {e}') from e

Prevention

When it happens

Trigger: Loading a RIFE .pth into the frame-interpolation node (comfy_extras/nodes_frame_interpolation.py calls detect_rife_config(sd) before constructing IFNet) where the checkpoint is an older/newer RIFE version whose block keys are absent or shaped differently — e.g. RIFE v2 checkpoints, community 'rife47' variants with renumbered blocks, or pruned/re-keyed state dicts.

Common situations: Downloading RIFE weights from community mirrors (flownet.pkl / .pth from HF) that are a different version than the supported 4.x family; renaming a trained/finetuned checkpoint; or using checkpoint-conversion tooling that strips 'module.' prefixes or renumbers blocks, breaking the key pattern the detector expects.

Related errors

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/4f7c536493a06500. Report an issue: GitHub.