Comfy-Org/ComfyUI · error · ValueError

Unknown rope_img: {rope_img}

Error message

Unknown rope_img: {rope_img}

What it means

calc_sizes accepts only two rope_img modes: 'extend' (plain grid over (th, tw)) and 'baseXXX' (interpolate relative to a base resolution, where XXX is the pixel size divided by 8*patch_size). Any other string is rejected because there is no defined way to build the position grid.

Source

Thrown at comfy/ldm/hydit/posemb_layers.py:205

        freqs_sin = freqs.sin().repeat_interleave(2, dim=1)  # [S, D]
        return freqs_cos, freqs_sin
    else:
        freqs_cis = torch.polar(torch.ones_like(freqs), freqs)  # complex64     # [S, D/2]
        return freqs_cis



def calc_sizes(rope_img, patch_size, th, tw):
    if rope_img == 'extend':
        # Expansion mode
        sub_args = [(th, tw)]
    elif rope_img.startswith('base'):
        # Based on the specified dimensions, other dimensions are obtained through interpolation.
        base_size = int(rope_img[4:]) // 8 // patch_size
        start, stop = get_fill_resize_and_crop((th, tw), base_size)
        sub_args = [start, stop, (th, tw)]
    else:
        raise ValueError(f"Unknown rope_img: {rope_img}")
    return sub_args


def init_image_posemb(rope_img,
                      resolutions,
                      patch_size,
                      hidden_size,
                      num_heads,
                      log_fn,
                      rope_real=True,
                      ):
    freqs_cis_img = {}
    for reso in resolutions:
        th, tw = reso.height // 8 // patch_size, reso.width // 8 // patch_size
        sub_args = calc_sizes(rope_img, patch_size, th, tw)
        freqs_cis_img[str(reso)] = get_2d_rotary_pos_embed(hidden_size // num_heads, *sub_args, use_real=rope_real)
        log_fn(f"    Using image RoPE ({rope_img}) ({'real' if rope_real else 'complex'}): {sub_args} | ({reso}) "
               f"{freqs_cis_img[str(reso)][0].shape if rope_real else freqs_cis_img[str(reso)].shape}")

View on GitHub (pinned to 1c6d8d45b3)

Solutions

1. Use 'extend' for plain grids, or 'base' + integer size like 'base1024' for the interpolated mode
2. Check the model's config file for the exact rope_img string the checkpoint was trained with
3. If you need a new mode, add an explicit branch rather than passing an unknown string

Example fix

# before
rope_img = "1024"  # ValueError
# after
rope_img = "base1024"

assert rope_img == 'extend' or (rope_img.startswith('base') and rope_img[4:].isdigit()), rope_img

def is_valid_rope_img(v: str) -> bool:
    return v == 'extend' or (v.startswith('base') and v[4:].isdigit())

When it happens

Trigger: Calling init_image_posemb/calc_sizes with rope_img values like 'xy', 'base', '1024', or a typo such as 'extenda'; the 'base' branch requires trailing digits (e.g. 'base512').

Common situations: Ported config keys from other DiT implementations (Flux/Wan use different rope names), typos, or checkpoint configs with unsupported bucketing strategies.

Related errors

• Got {params.axes_dim} but expected positional dim {pe_dim}
• Got {params.axes_dim} but expected positional dim {pe_dim}
• Unknown interpolation method {self.interpolation}
• Got {params.axes_dim} but expected positional dim {pe_dim}
• Got {params.axes_dim} but expected positional dim {pe_dim}

AI-assisted analysis of Comfy-Org/ComfyUI@1c6d8d45b3 (2026-08-14). Data as JSON: /api/errors/e2686e8cdf667579. Report an issue: GitHub.