Ace-Step-1.5-ScragVAE

[jt858766-cmyk]

Custom Node Testing

• I have tried disabling custom nodes and the issue persists (see how to disable custom nodes if you need help)

Expected Behavior

Ace-Step-1.5-ScragVAE not working

Actual Behavior

Steps to Reproduce

Debug Logs

Output will be ignored
WARNING: No VAE weights detected, VAE not initalized.
!!! Exception during processing !!! ERROR: VAE is invalid: None

If the VAE is from a checkpoint loader node your checkpoint does not contain a valid VAE.
Traceback (most recent call last):
  File "D:\ComfyUI_2026.04.21\ComfyUI\execution.py", line 534, in execute
    output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
                                                              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\ComfyUI_2026.04.21\ComfyUI\execution.py", line 334, in get_output_data
    return_values = await _async_map_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "D:\ComfyUI_2026.04.21\ComfyUI\execution.py", line 308, in _async_map_node_over_list
    await process_inputs(input_dict, i)
  File "D:\ComfyUI_2026.04.21\ComfyUI\execution.py", line 296, in process_inputs
    result = f(**inputs)
  File "D:\ComfyUI_2026.04.21\ComfyUI\nodes.py", line 831, in load_vae
    vae.throw_exception_if_invalid()
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
  File "D:\ComfyUI_2026.04.21\ComfyUI\comfy\sd.py", line 869, in throw_exception_if_invalid
    raise RuntimeError("ERROR: VAE is invalid: None\n\nIf the VAE is from a checkpoint loader node your checkpoint does not contain a valid VAE.")
RuntimeError: ERROR: VAE is invalid: None

If the VAE is from a checkpoint loader node your checkpoint does not contain a valid VAE.

Other

No response

[comfyanonymous]

It needs to be converted to the comfyui format first.

[romerojunior]

what a useless comment :) HOW????

[jovan2009]

what a useless comment :) HOW????

I went last night until about 4 AM on a side quest about the "How" to do this and I came up with nothing concrete, yet. My impression is that ComfyUI expects VAE weights to be named in a specific way. Inspecting Scrag VAE safetensors I see that the weights differ from original Ace-Step-1.5 VAE. So my working hypothesis is that the weights have to be renamed or something like that, possibly using tools from this node https://github.com/silveroxides/ComfyUI-ModelUtils.
I don't know, I am both ignorant in the matter and sleep deprived. I will continue when I am more able to do it. I am really curious how my Ace-step workflows would sound with Scrag VAE.

[jovan2009]

I did it!

Actually Gemma 4 E4B running in llama-server did it, I only explained the problem and came up with ideas on the way.

So the script does the job. I was able to run successfully one of my Ace-Step workflows with the converted ScragVAE.

You will need:

>python Convert_Scrag_VAE_01.01.00.py --help
usage: Convert_Scrag_VAE_01.01.00.py [-h] --source SOURCE --target TARGET --output OUTPUT

VAE Weight Migration Utility for Stable Diffusion/ComfyUI.

options:
  -h, --help       show this help message and exit
  --source SOURCE  Path to the finetuned VAE safetensors file (e.g., ScragVAE_FP32.safetensors).
  --target TARGET  Path to the original, working VAE safetensors file (e.g., ACE-1.5-VAE-checkpoint.safetensors).
  --output OUTPUT  Path where the new, migrated safetensors file will be saved.

--source = the unmodified ScragVAE file
--target = the original Ace-Step 1.5 VAE that works currently with ComfyUI, from which the script will try to replicate the structure to ScragVAE.
--output = path to the new converted ScragVAE that will be generated by running the script

Here is the script directly from Gemma 4 mouth :)))

Convert_Scrag_VAE_01.01.00.py

I'm not a developer, use it at your own risk. Actually, before running it, I made a backup to the safetensors files involved in case it decides to overwrite them or something. It wasn't the case but you never know, if you didn't made the code yourself you have to be careful.

Scratch that, WIP. In the current form the script actually just makes a file copy of the current "official" Ace-Step 1.5 VAE instead a "conversion" of the ScragVAE. I was bamboozled by Gemma, LOL.

[fappaz]

I vibecoded this script to convert it to ComfyUI format. Not sure if it's correct because the difference is very subtle, but it doesn't throw any errors in the console.

import re
import torch
from safetensors.torch import load_file, save_file

# Path to your downloaded ScragVAE file (https://huggingface.co/scragnog/Ace-Step-1.5-ScragVAE/blob/main/diffusion_pytorch_model.safetensors)
input_file = "diffusion_pytorch_model.safetensors"
output_file = "Ace-Step-1.5-ScragVAE-COMFYUI.safetensors"

# ── Architecture mapping ──────────────────────────────────────────────────────
#
# ScragVAE uses named-module keys; ComfyUI wraps the same model in
# nn.Sequential so every sub-layer gets a numeric index.
#
# ENCODER
#   encoder.conv1.*                              → encoder.layers.0.*
#   encoder.block.N.res_unitX.snake1.*           → encoder.layers.(N+1).layers.(X-1).layers.0.*   (+ squeeze [1,C,1]→[C])
#   encoder.block.N.res_unitX.conv1.*            → encoder.layers.(N+1).layers.(X-1).layers.1.*
#   encoder.block.N.res_unitX.snake2.*           → encoder.layers.(N+1).layers.(X-1).layers.2.*   (+ squeeze)
#   encoder.block.N.res_unitX.conv2.*            → encoder.layers.(N+1).layers.(X-1).layers.3.*
#   encoder.block.N.snake1.*                     → encoder.layers.(N+1).layers.3.*                 (+ squeeze)
#   encoder.block.N.conv1.*                      → encoder.layers.(N+1).layers.4.*
#   encoder.snake1.*                             → encoder.layers.6.*                              (+ squeeze)
#   encoder.conv2.*                              → encoder.layers.7.*
#
# DECODER
#   decoder.conv1.*                              → decoder.layers.0.*
#   decoder.block.N.snake1.*                     → decoder.layers.(N+1).layers.0.*                 (+ squeeze)
#   decoder.block.N.conv_t1.*                    → decoder.layers.(N+1).layers.1.*
#   decoder.block.N.res_unitX.snake1.*           → decoder.layers.(N+1).layers.(X+1).layers.0.*   (+ squeeze)
#   decoder.block.N.res_unitX.conv1.*            → decoder.layers.(N+1).layers.(X+1).layers.1.*
#   decoder.block.N.res_unitX.snake2.*           → decoder.layers.(N+1).layers.(X+1).layers.2.*   (+ squeeze)
#   decoder.block.N.res_unitX.conv2.*            → decoder.layers.(N+1).layers.(X+1).layers.3.*
#   decoder.snake1.*                             → decoder.layers.6.*                              (+ squeeze)
#   decoder.conv2.*                              → decoder.layers.7.*
#
# Snake activations are stored as [1, C, 1] in ScragVAE but [C] in ComfyUI.
# ─────────────────────────────────────────────────────────────────────────────

# Patterns ordered most-specific first so earlier matches win.
# Each entry: (compiled_pattern, replacement_template, needs_squeeze)
PATTERNS = [
    # ── encoder ──
    # encoder.block.N.res_unitX.snake{1,2}.param
    (re.compile(r'^encoder\.block\.(\d+)\.res_unit(\d+)\.snake(\d+)\.(alpha|beta)$'),
     lambda m: f"encoder.layers.{int(m.group(1))+1}.layers.{int(m.group(2))-1}.layers.{(int(m.group(3))-1)*2}.{m.group(4)}",
     True),
    # encoder.block.N.res_unitX.conv{1,2}.param
    (re.compile(r'^encoder\.block\.(\d+)\.res_unit(\d+)\.conv(\d+)\.(bias|weight_g|weight_v)$'),
     lambda m: f"encoder.layers.{int(m.group(1))+1}.layers.{int(m.group(2))-1}.layers.{(int(m.group(3))-1)*2+1}.{m.group(4)}",
     False),
    # encoder.block.N.snake1.param  (block-level snake before downsampler)
    (re.compile(r'^encoder\.block\.(\d+)\.snake1\.(alpha|beta)$'),
     lambda m: f"encoder.layers.{int(m.group(1))+1}.layers.3.{m.group(2)}",
     True),
    # encoder.block.N.conv1.param  (block downsampler)
    (re.compile(r'^encoder\.block\.(\d+)\.conv1\.(bias|weight_g|weight_v)$'),
     lambda m: f"encoder.layers.{int(m.group(1))+1}.layers.4.{m.group(2)}",
     False),
    # encoder.conv1.*  (initial conv)
    (re.compile(r'^encoder\.conv1\.(bias|weight_g|weight_v)$'),
     lambda m: f"encoder.layers.0.{m.group(1)}",
     False),
    # encoder.snake1.*  (final snake)
    (re.compile(r'^encoder\.snake1\.(alpha|beta)$'),
     lambda m: f"encoder.layers.6.{m.group(1)}",
     True),
    # encoder.conv2.*  (final conv to latent)
    (re.compile(r'^encoder\.conv2\.(bias|weight_g|weight_v)$'),
     lambda m: f"encoder.layers.7.{m.group(1)}",
     False),

    # ── decoder ──
    # decoder.block.N.res_unitX.snake{1,2}.param
    (re.compile(r'^decoder\.block\.(\d+)\.res_unit(\d+)\.snake(\d+)\.(alpha|beta)$'),
     lambda m: f"decoder.layers.{int(m.group(1))+1}.layers.{int(m.group(2))+1}.layers.{(int(m.group(3))-1)*2}.{m.group(4)}",
     True),
    # decoder.block.N.res_unitX.conv{1,2}.param
    (re.compile(r'^decoder\.block\.(\d+)\.res_unit(\d+)\.conv(\d+)\.(bias|weight_g|weight_v)$'),
     lambda m: f"decoder.layers.{int(m.group(1))+1}.layers.{int(m.group(2))+1}.layers.{(int(m.group(3))-1)*2+1}.{m.group(4)}",
     False),
    # decoder.block.N.snake1.param  (block-level snake before upsampler)
    (re.compile(r'^decoder\.block\.(\d+)\.snake1\.(alpha|beta)$'),
     lambda m: f"decoder.layers.{int(m.group(1))+1}.layers.0.{m.group(2)}",
     True),
    # decoder.block.N.conv_t1.param  (transposed conv upsampler)
    (re.compile(r'^decoder\.block\.(\d+)\.conv_t1\.(bias|weight_g|weight_v)$'),
     lambda m: f"decoder.layers.{int(m.group(1))+1}.layers.1.{m.group(2)}",
     False),
    # decoder.conv1.*  (initial decoder conv)
    (re.compile(r'^decoder\.conv1\.(bias|weight_g|weight_v)$'),
     lambda m: f"decoder.layers.0.{m.group(1)}",
     False),
    # decoder.snake1.*  (final snake)
    (re.compile(r'^decoder\.snake1\.(alpha|beta)$'),
     lambda m: f"decoder.layers.6.{m.group(1)}",
     True),
    # decoder.conv2.*  (final output conv, no bias)
    (re.compile(r'^decoder\.conv2\.(weight_g|weight_v)$'),
     lambda m: f"decoder.layers.7.{m.group(1)}",
     False),
]


def convert_key(key, tensor):
    for pattern, replacement, needs_squeeze in PATTERNS:
        m = pattern.match(key)
        if m:
            new_key = replacement(m)
            new_tensor = tensor.squeeze() if needs_squeeze else tensor
            return new_key, new_tensor
    return None, None


state_dict = load_file(input_file)
new_state_dict = {}
unmatched = []

for key, value in state_dict.items():
    new_key, new_value = convert_key(key, value)
    if new_key is None:
        unmatched.append(key)
    else:
        new_state_dict[new_key] = new_value

if unmatched:
    print(f"WARNING: {len(unmatched)} keys could not be mapped:")
    for k in unmatched:
        print(f"  {k}")
else:
    print(f"All {len(state_dict)} keys mapped successfully.")

save_file(new_state_dict, output_file)
print(f"Saved '{output_file}'. Use it in the ComfyUI VAE Loader node.")

Link to the VAE file already converted: https://civitai.red/models/2425543?modelVersionId=2907969

Sample workflow:
ace15-metal.json

Comparison

Standard ACE-Step 1.5 VAE:
metal-ace-standard-vae.mp3

ScragVAE VAE:
metal-ace-scrag-vae.mp3