Replies: 5 comments 27 replies
|
I've installed it through the script as above, but the custom node fails to import. The script created the folder with all the files as expected, no errors during installation. There is nothing in the startup logs to understand why it can't be imported, so I can't say what problem might be |
|
I don't think we have a different directory structure, I have ComfyUI Portable. I ran the command using Git Bash on Windows: 
As you can see the script installed succesfully the node in the custom node folder, but when I run ComfyUI the custom node can't be imported. After running with all nodes disabled except yours, I noticed this warning message in the logs: [WARNING] Skip E:\ComfyUI\ComfyUI\custom_nodes\infinity-diffusion module for custom nodes due to the lack of NODE_CLASS_MAPPINGS or comfy_entrypoint (need one).Here's the full startup logs:E:\ComfyUI>.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --disable-all-custom-nodes --whitelist-custom-nodes infinity-diffusion
[INFO] setup plugin alembic.autogenerate.schemas
[INFO] setup plugin alembic.autogenerate.tables
[INFO] setup plugin alembic.autogenerate.types
[INFO] setup plugin alembic.autogenerate.constraints
[INFO] setup plugin alembic.autogenerate.defaults
[INFO] setup plugin alembic.autogenerate.comments
[INFO] Prestartup Skipping comfyui-manager due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Prestartup Skipping ComfyUI_Viewer due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Prestartup Skipping rgthree-comfy due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Found comfy_kitchen backend eager: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['adaln', 'apply_rope', 'apply_rope1', 'apply_rope_split_half', 'apply_rope_split_half1', 'convrot_w4a4_linear', 'dequantize_convrot_w4a4_weight', 'dequantize_int8_convrot_weight', 'dequantize_int8_convrot_weight_dtype', 'dequantize_int8_simple', 'dequantize_int8_simple_dtype', 'dequantize_mxfp8', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'gemv_awq_w4a16', 'int8_linear', 'prepare_int4_weight_for_int8_linear', 'quantize_and_rotate_rowwise', 'quantize_convrot_w4a4_weight', 'quantize_int8_convrot_weight', 'quantize_int8_rowwise', 'quantize_int8_tensorwise', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'quantize_svdquant_w4a4', 'rms_rope', 'rms_rope1', 'rms_rope_split_half', 'rms_rope_split_half1', 'scaled_mm_mxfp8', 'scaled_mm_nvfp4', 'scaled_mm_svdquant_w4a4', 'stochastic_rounding_fp8']}
[INFO] Found comfy_kitchen backend cuda: {'available': True, 'disabled': False, 'unavailable_reason': None, 'capabilities': ['adaln', 'apply_rope', 'apply_rope1', 'apply_rope_split_half', 'apply_rope_split_half1', 'convrot_w4a4_linear', 'dequantize_convrot_w4a4_weight', 'dequantize_int8_convrot_weight', 'dequantize_int8_convrot_weight_dtype', 'dequantize_int8_simple', 'dequantize_int8_simple_dtype', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'gemv_awq_w4a16', 'int8_linear', 'prepare_int4_weight_for_int8_linear', 'quantize_and_rotate_rowwise', 'quantize_convrot_w4a4_weight', 'quantize_int8_convrot_weight', 'quantize_int8_rowwise', 'quantize_int8_tensorwise', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8', 'quantize_svdquant_w4a4', 'rms_rope', 'rms_rope1', 'rms_rope_split_half', 'rms_rope_split_half1', 'scaled_mm_nvfp4', 'scaled_mm_svdquant_w4a4', 'stochastic_rounding_fp8']}
[INFO] Found comfy_kitchen backend triton: {'available': True, 'disabled': True, 'unavailable_reason': None, 'capabilities': ['adaln', 'apply_rope', 'apply_rope1', 'apply_rope_split_half', 'apply_rope_split_half1', 'dequantize_nvfp4', 'dequantize_per_tensor_fp8', 'int8_linear', 'quantize_and_rotate_rowwise', 'quantize_int8_rowwise', 'quantize_mxfp8', 'quantize_nvfp4', 'quantize_per_tensor_fp8']}
[INFO] Checkpoint files will always be loaded safely.
[INFO] Total VRAM 8192 MB, total RAM 65277 MB
[INFO] pytorch version: 2.13.0+cu130
[INFO] Set vram state to: NORMAL_VRAM
[INFO] Device: cuda:0 NVIDIA GeForce RTX 3070 Ti Laptop GPU : cudaMallocAsync
[INFO] Using async weight offloading with 2 streams
[INFO] Enabled pinned memory 26110.0
[INFO] Using pytorch attention
aimdo: src-win/cuda-detour.c:38:INFO:aimdo_setup_hooks: installing 6 hooks
aimdo: src-win/cuda-detour.c:28:DEBUG:install_hook_entries: hooks successfully installed
aimdo: src-win/shmem-detect.c:80:INFO:comfy-aimdo WDDM adapter match: NVIDIA GeForce RTX 3070 Ti Laptop GPU runtime_luid=00000000:00011eef dxgi_luid=00000000:00011eef
aimdo: src/control.c:248:INFO:comfy-aimdo inited for GPU: NVIDIA GeForce RTX 3070 Ti Laptop GPU (VRAM: 8191 MB)
[INFO] DynamicVRAM support detected and enabled
[INFO] Python version: 3.12.7 (tags/v3.12.7:0b05ead, Oct 1 2024, 03:06:41) [MSC v.1941 64 bit (AMD64)]
[INFO] ComfyUI version: 0.28.0
[INFO] comfy-aimdo version: 0.4.10
[INFO] comfy-kitchen version: 0.2.22
[INFO] comfyui-frontend-package version: 1.45.21
[INFO] comfyui-workflow-templates version: 0.11.12
[INFO] comfyui-embedded-docs version: 0.5.8
[INFO] comfy-kitchen version: 0.2.22
[INFO] comfy-aimdo version: 0.4.10
[INFO] [Prompt Server] web root: E:\ComfyUI\python_embeded\Lib\site-packages\comfyui_frontend_package\static
[INFO] Asset seeder disabled
[INFO] No OpenGL_accelerate module loaded: Acceleration disabled
[INFO] Skipping ComfyUI-bleh due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping comfyui-custom-scripts due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping ComfyUI-DCW due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping ComfyUI-DeGrid due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping ComfyUI-Impact-Pack due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping comfyui-impact-subpack due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping comfyui-inspire-pack due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping comfyui-kjnodes due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping comfyui-lora-loader due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping comfyui-lora-optimizer due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping comfyui-manager due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping ComfyUI-MemoryVisualization due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping ComfyUI-NAG-Extended due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping ComfyUI-Optical-Realism due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping Comfyui-Resolution-Master due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping ComfyUI-Universal-Seamless-Tiles due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping ComfyUI-UtilsCollection due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping comfyui-videohelpersuite due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping comfyui_essentials due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping ComfyUI_Fens_Simple_Nodes due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping ComfyUI_LMStudio_EasyQuery due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping comfyui_nvidia_rtx_nodes due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping comfyui_randomperson due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping ComfyUI_Viewer due to disable_all_custom_nodes and whitelist_custom_nodes
# Registered infinity sampler and scheduler
[WARNING] Skip E:\ComfyUI\ComfyUI\custom_nodes\infinity-diffusion module for custom nodes due to the lack of NODE_CLASS_MAPPINGS or comfy_entrypoint (need one).
[INFO] Skipping loadImageWithSubfolders.py due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping rgthree-comfy due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping SesquiLSR due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping was-ns due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO] Skipping websocket_image_save.py due to disable_all_custom_nodes and whitelist_custom_nodes
[INFO]
Import times for custom nodes:
[INFO] 0.0 seconds (IMPORT FAILED): E:\ComfyUI\ComfyUI\custom_nodes\infinity-diffusion
[INFO]
[INFO] Context impl SQLiteImpl.
[INFO] Will assume non-transactional DDL.
[INFO] Using RAM pressure cache.
[INFO] Starting server
[INFO] To see the GUI go to: http://127.0.0.1:8188 |
|
I don't see anything wrong with your logs. In the process of making the script, I tried that and you should see infinity as an option for the KSampler node, both for sampler and scheduler dropdowns. If that doesn't work, I might try comparing the script against my ComfyUI directory again to see if I did miss anything.
It's not supposed to be imported. You will find infinity as an option in both sampler and scheduler dropdowns for KSampler node. If your config already uses a KSampler node, the script will seamlessly add/remove infinity to/from the node. Okay I'll try checking this part and inform you if I found something: |
|
Okay I pushed a commit (galpt/infinity-diffusion@7e8d712) and it seems I did miss something that's causing your setup to not import the custom node. |
Ok now it's imported correctly! 🎉 
|
|
@LukeG89 |
|
@galpt So far I found 3 issues:
This is the error I get:[ERROR] !!! Exception during processing !!! last element of sigmas must be 0
[ERROR] Traceback (most recent call last):
File "E:\ComfyUI\ComfyUI\execution.py", line 543, 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 "E:\ComfyUI\ComfyUI\execution.py", line 342, 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 "E:\ComfyUI\ComfyUI\execution.py", line 316, in _async_map_node_over_list
await process_inputs(input_dict, i)
File "E:\ComfyUI\ComfyUI\execution.py", line 304, in process_inputs
result = f(**inputs)
^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\nodes.py", line 1641, in sample
return common_ksampler(model, noise_seed, steps, cfg, sampler_name, scheduler, positive, negative, latent_image, denoise=denoise, disable_noise=disable_noise, start_step=start_at_step, last_step=end_at_step, force_full_denoise=force_full_denoise)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\nodes.py", line 1571, in common_ksampler
samples = comfy.sample.sample(model, noise, steps, cfg, sampler_name, scheduler, positive, negative, latent_image,
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\comfy\sample.py", line 74, in sample
samples = sampler.sample(noise, positive, negative, cfg=cfg, latent_image=latent_image, start_step=start_step, last_step=last_step, force_full_denoise=force_full_denoise, denoise_mask=noise_mask, sigmas=sigmas, callback=callback, disable_pbar=disable_pbar, seed=seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\comfy\samplers.py", line 1444, in sample
return sample(self.model, noise, positive, negative, cfg, self.device, sampler, sigmas, self.model_options, latent_image=latent_image, denoise_mask=denoise_mask, callback=callback, disable_pbar=disable_pbar, seed=seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\comfy\samplers.py", line 1334, in sample
return cfg_guider.sample(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\comfy\samplers.py", line 1316, in sample
output = executor.execute(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\comfy\patcher_extension.py", line 113, in execute
return self.original(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\comfy\samplers.py", line 1254, in outer_sample
output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\comfy\samplers.py", line 1229, in inner_sample
samples = executor.execute(self, sigmas, extra_args, callback, noise, latent_image, denoise_mask, disable_pbar)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\comfy\patcher_extension.py", line 113, in execute
return self.original(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\comfy\samplers.py", line 999, in sample
samples = self.sampler_function(model_k, noise, sigmas, extra_args=extra_args, callback=k_callback, disable=disable_pbar, **self.extra_options)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\python_embeded\Lib\site-packages\torch\utils\_contextlib.py", line 124, in decorate_context
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\custom_nodes\infinity-diffusion\infinity_comfyui\integration.py", line 54, in sample_infinity
return sampler.sample(denoise_fn, x, sigmas, callback=callback)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\python_embeded\Lib\site-packages\torch\utils\_contextlib.py", line 124, in decorate_context
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\custom_nodes\infinity-diffusion\infinity_diffusion.py", line 223, in sample
raise ValueError("last element of sigmas must be 0")
ValueError: last element of sigmas must be 0I wanted to try your infinite sampler+scheduler in my Krea2 workflow, but I've several custom stuff going on there, like starting the generation with Raw model and then continues with Turbo, but doesn't work for me due to these problems. Here's a test workflow I made to find the possible cause of the 2nd and 3rd issues: |
|
So the sampler+scheduler combo always works on my end, but the setup I use consists of base image generation → upscale (latent) → face detailer → eyes detailer. I think it is still considered basic compared to your custom setup, so that is probably why mine always works. If our directory structure and the files inside it are the same, the errors are probably related to your custom setup. |
|
I pushed galpt/infinity-diffusion@cb13ced that might fix your issue.
That is intentional. The fix in the new commit should prevent certain models from triggering the infinite self-correction loop. |
Tried both NameError: name 'inv_reversed' is not defined[ERROR] !!! Exception during processing !!! name 'inv_reversed' is not defined
[ERROR] Traceback (most recent call last):
File "E:\ComfyUI\ComfyUI\execution.py", line 543, 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 "E:\ComfyUI\ComfyUI\execution.py", line 342, 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 "E:\ComfyUI\ComfyUI\execution.py", line 316, in _async_map_node_over_list
await process_inputs(input_dict, i)
File "E:\ComfyUI\ComfyUI\execution.py", line 304, in process_inputs
result = f(**inputs)
^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\comfy_api\internal\__init__.py", line 149, in wrapped_func
return method(locked_class, **inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\comfy_api\latest\_io.py", line 1935, in EXECUTE_NORMALIZED
to_return = cls.execute(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\comfy_extras\nodes_custom_sampler.py", line 1048, in execute
samples = guider.sample(noise.generate_noise(latent), latent_image, sampler, sigmas, denoise_mask=noise_mask, callback=callback, disable_pbar=disable_pbar, seed=noise.seed)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\comfy\samplers.py", line 1316, in sample
output = executor.execute(noise, latent_image, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\comfy\patcher_extension.py", line 113, in execute
return self.original(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\comfy\samplers.py", line 1254, in outer_sample
output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed, latent_shapes=latent_shapes)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\comfy\samplers.py", line 1229, in inner_sample
samples = executor.execute(self, sigmas, extra_args, callback, noise, latent_image, denoise_mask, disable_pbar)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\comfy\patcher_extension.py", line 113, in execute
return self.original(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\comfy\samplers.py", line 999, in sample
samples = self.sampler_function(model_k, noise, sigmas, extra_args=extra_args, callback=k_callback, disable=disable_pbar, **self.extra_options)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\python_embeded\Lib\site-packages\torch\utils\_contextlib.py", line 124, in decorate_context
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\custom_nodes\infinity-diffusion\infinity_comfyui\integration.py", line 59, in sample_infinity
return sampler.sample(denoise_fn, x, sigmas, callback=callback)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\python_embeded\Lib\site-packages\torch\utils\_contextlib.py", line 124, in decorate_context
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI\ComfyUI\custom_nodes\infinity-diffusion\infinity_diffusion.py", line 295, in sample
elif inv_reversed:
^^^^^^^^^^^^
NameError: name 'inv_reversed' is not defined
Ok don't worry, you can use the same sampler setups from my test workflow with other models if you want to test compatibility. Krea2 is the latest, very popular and powerful open-source model available at the moment, so if infinity works with it would be great. I don't know how much RAM you have, but with int8 models you should be able to run Krea2 thanks to the offloading (memory management nowadays is super efficient). qwen3vl_4b_int8.safetensors is a lighter TE for it. |
|
Thanks again for the detailed reports! The NameError (inv_reversed not defined) was from an earlier build of the main/research branches. The current realism branch has a completely rewritten sampler that doesn't use that variable at all, so it should be fine now. I've tested it with basic workflows and it works, but I'd love to hear if it runs into anything on your end. 64GB DDR5 should be plenty for Krea2, especially with the lighter text encoder you mentioned. I'm intrigued to try it. The offloading in newer ComfyUI builds is really impressive with what it can handle on mid-range VRAM. Thanks for the link to the int8 TE, that's helpful. Really appreciate you taking the time to test across different configs! |
|
faka you |
|
I created a new branch here: This one aims to be a good alternative to DPM++ 2M Karras, which most people say is good for realistic images. The self-correcting logic might be beneficial for use cases that DPM++ 2M Karras does not cover. Infinity is meant to be used together as both the sampler and scheduler, so while you can use for example Euler A + Infinity scheduler, you do not get the self-correcting logic of Infinity. The self-correction is a sampler feature that happens to modify the sigma schedule at runtime. Making it work independently of the infinity sampler would require moving the invariant-tracking and insertion logic into a shared wrapper that any sampler could use, which might require making modifications to different areas in ComfyUI. |
It's working with both single and dual sampler configurations, but seems like it's not really compatible with Krea2 Turbo. Also I get two different results, even though both configurations should output the same image since they use identical settings. With a single sampler, the output is too noisy: And with dual sampler, it's oversaturated: The noisy output is probably because Krea2 doesn't like staying too much in the low sigmas (final 7-8 steps): 
For reference, this is 
By the way, there are other samplers that can be considered instead of DPM++ 2M Karras. For example Not to mention the RES samplers implemented by RES4LYF, especially I don't know if this can be useful for you, but you can also consider Colored Noise Diffusion Sampling for the sde/ancestral samplers, there is a wrapper node for it: ComfyUI-ColoredNoiseDiffusionSampling |
|
Ok, I pulled the latest commit (galpt/infinity-diffusion@78af7ba) and ran a quick test. Now the noisy output is gone with the single sampler configuration, no more artifacts! 🎉 But the output is slightly contrasted and saturated compared to dpmpp_2m. The oversaturation problem with the dual sampler configuration still persists. Here's the comparisons: 
I tried to understand if the problem was the sigmas split, but apparently it's not. This is a comparison between 
So I guess that the oversaturation is caused by the infinity sampler for some reason. It happens in the second sampler. infinity_dual_sampler_test.mp4 |
|
Glad the artifacts are gone, that's good to hear. The contrast difference vs DPM++ 2M is expected. The variance stabiliser adjusts per-channel standard deviation to preserve volumetric depth, which is fundamentally different from DPM++ 2M's edge-amplifying second-order correction. Different math produces different results, and that's the whole point of having multiple samplers to choose from. Expecting them to produce the same output is like expecting two different formulas to generate the same apple. The dual sampler thing isn't a bug. It affects any sampler that tracks history across steps. The EMA resets every time sample() is called, and reusing it across stages wouldn't make sense because the noise level distribution is different in each stage. DPM++ 2M has the same limitation with its own history (I checked the source), it just isn't as noticeable because its correction is weaker. If you check the Visual comparison section in the README, I used 20 steps at CFG 6.0 with a single sampler and didn't get oversaturation. Following that exact setup (step count, CFG, single sampler) should let you reproduce the same behaviour. Infinity is designed for single-pass sampling at a reasonable step count. Things like dual sampler or too few steps will easily go wrong, but that's how the formulas work, not a bug. Compared side by side with DPM++ 2M or Euler under those conditions it should behave similar or better. |
|
@LukeG89 okay so I'm trying to setup your krea2 config now, currently in the process of downloading the files. |
|
@LukeG89 I tested it and I plan to stop here for Krea2. Turbo works differently and the Raw is too slow for my GPU so I gave up. 
As you can see, the normal node does not generate an image that looks really bad at all. It is just that some people are probably already used to how DPM or Euler or any samplers they usually work with look like, and the brain is telling you Infinity is wrong, while I do not think so because it is simply using a different math, so different colors and details will be visible. Tested using the nano branch. This branch is using formulas different from Euler or DPM so what you are seeing is actually the output of the math. If we ask the math to produce the same result as DPM or Euler, it is not fair and is similar to asking 2 different math formulas to give the same taste, which is not how math works. So I think Infinity will still give users a new alternative when they want to try something new. For SDXL, all combinations at 896x1152 portrait, 20 steps, CFG 6.0, seed 210895200085864, same model (waiMatureIllustrious v2.0, SDXL). This is from the realism branch:
and this is for the new 
I will focus on the nano branch from now on since it is an original compared to other branches which are using formulas like Euler or DPM++ Solver which are probably already used by other samplers in ComfyUI. Need more people to truly test it with their own different unique use cases because my workflow and your Krea2 Turbo workflow might not be enough to explore what the new math formulas are capable of and which parts should be optimized. |
|
Honestly, the idea of a sampler that remembers the direction of change over the last few steps is intriguing. Most samplers I've worked with, like The fact that you're checking for overshooting and scaling down the correction when needed is a good safeguard. I've seen similar techniques used in some RLHF (Reinforcement Learning from Human Feedback) implementations to stabilize the training process. One thing that caught my eye is the scheduler shifting step budget toward the end. I've experimented with similar step reallocation techniques when fine-tuning models with For example, in my own code, I've used a simple reallocation function like this: def reallocate_steps(steps, ratio=0.2):
initial_steps = int(steps * (1 - ratio))
final_steps = steps - initial_steps
return initial_steps, final_stepsThis reallocation can help the model refine details in the later stages. Curious to hear if you've experimented with different reallocation ratios or if there's a particular reason you chose to shift budget toward the end. Would love to know how this affects the overall quality and if you've noticed any trade-offs. |
|
Thanks for the interest and the thoughtful feedback! The step reallocation idea with LoRA is genuinely something I hadn't considered before, and it's the first time I've heard about it from a user. LukeG89's reports about Krea2 compatibility also gave me some ideas. Your reallocate_steps function shares the same intuition as the sine perturbation in the infinity scheduler, shifting budget toward the end for final cleanup. The sine formula f(u) = u - ssin(piu)/pi is nice because it's a smooth deformation of the whole distribution rather than a hard split. I've been thinking about whether combining a hard reallocation (like your function) with a smooth sinusoid could let users select profiles, but that goes against the no-knobs philosophy. I'll try to see if I can fit the idea into the current Infinity design. Might make a new branch specifically for this experiment. |
|
I like the sampler overall, but the installation was difficult because the provided .sh file didn't accept my path, so I had to edit it manually. Performance is much better, reducing my required steps from 300–400 to around 100. I miss the console progress output, but the progress bar in the main window is sufficient. On my Windows 10 system (RX 9060 XT 16 GB, Ryzen 5 5500, 32 GB RAM), generation is fast, but generating 10–12 images at once causes brief system freezes (micro stutters) every 5–8 seconds. I'm not sure if generating one image per 10 runs would avoid this issue. I hope you improve the Installation process in the future and the console progress bar. |
|
Hi @blepping, Thanks for the detailed feedback. I admit it was made with the help of an LLM. You seem very knowledgeable in this field, and I can indeed be considered new to it, so I hope you don't mind helping shape where this project should go. We could create a new branch based on what you think the correct approach is. The ETA map approach for the noise injection and the idea of training a small classifier per latent format both sound like the kind of direction I need, so I'd like to start from there. I also do research in other fields, like kernel schedulers and more, if you check my other projects. I've learned that if we don't understand something, we shouldn't assume it doesn't make sense yet. It is almost always the case that something makes sense after someone publishes a paper or a technical post about it, so I'm not limiting myself to saying "no, I won't try it" just because ComfyUI or other tools haven't tried it yet. Of course doing it blindly is bad, so if you have some knowledge to share, I'd gladly accept your suggestions on how the project should be shaped for the next branch. Please go easy on me. Since I'm a beginner, I'd appreciate constructive feedback and will gladly take your suggestions. The reason I made it open source and posted it here as a discussion is so that people more knowledgeable than I am can join in and help me shape the project toward the correctness it should have. I'm hoping that, given how many stars the ComfyUI repo has as of today, at least some of them would be willing to help. I would try my best to learn, that is, if people who understand this field don't mind sharing their wisdom and helping shape the implementation. Replies providing links and references so I know where to look would be greatly appreciated. I take the warning about relying on LLMs seriously, but I don't think a reply that only says it's bad is enough to help shape the project. If this project can be implemented correctly and proven to be useful, many people could benefit from it. That's the point of making it open source and discussing it here in the first place. |
That's a good attitude! I have a blunt communication style but just to be clear, I was not saying your project is bad, I don't want you to feel bad, give up, whatever. The main thing I'd criticize is making definite claims about something that (based on everything I know) couldn't be factually accurate and with vibe coding/etc becoming more popular this is happening more frequently. Also, when I mentioned reinventing the wheel, math canceling out, etc I wasn't necessarily saying any of that applies to your project (just in case you interpreted it that way), just that you can't really count on LLMs for a reality check. I can't claim to be "very knowledgeable" or an expert, I have been interested in this as a hobby for a while but there is a lot that I don't know. I don't want to take over your project and I don't really have the time to really take something under my wing as it were. I haven't had the time to maintain my own projects reliably for some time now.
When I said it doesn't make sense, that was based on understanding (or believing I understand) the thing. If you understand a subject, you can be fairly confident when you encounter something like a claim that you know just can't be possible. Based on what I know, your claims about material classification just can't be correct and from skimming the code I didn't see anything that changed my mind. It's not a good thing for you to believe that something is possible when it isn't, since I assume you want to live in reality and believe things that are true. If someone looks at your repo and reads "Every pixel is classified by material using Laws' texture-energy masks, then enhanced according to what it is" unless they have the technical knowledge to understand why it's not really possible to do that, they're going naturally going to think a sampler that can do that is going to be better (and it probably would be, if you actually could, though figuring out how to apply the enhancements would also be an issue). Someone else that's learning might come away thinking that doing that kind of classification on arbitrary latents is possible, or they look at the code and go "Ah ha, so that's how you do the classification". That's why I'm taking the time to criticize it, but that's just criticizing something (I believe) is not factually accurate, it's not a personal attack and I'm not implying you are or did anything intentionally dishonest. About the material classification, one immediate problem is that you are doing it per channel (as far as I can see, I did only skim the code). Are you getting the same values for your classification in every channel? Because if not, then it clearly doesn't make sense. If an area in the image is skin in channel 0, it doesn't make sense that it would be a different material in a different channel. Unlike a nice clean 3 channel RGB image, what a channel means in a latent is pretty much arbitrary and opaque. Another problem with trying to classify stuff in latent space is spatial compression. Even SDXL has 8x spatial compression and some models go higher (LTX is 32x, I believe). That means one element in your latent represents an 8x8 block of pixels in the image. It looks like maybe you mostly tested on SDXL (since your example in the repo is SDXL) and SDXL only has 4 channels. They're kind-of sort-of related to colors like a RGB image but once you get to something like a Flux latent with 16 channels, Flux 2 latent (64 channels if I remember correctly), LTX (128 channels) what "features" a channel represents are much less clear. If you're dealing with something like Qwen image or other models that use a 5D (video) latent then maybe some of the channels have something to do with temporality (even if it's only generating one frame). How do you analyze material on an arbitrary latent format, analyzing the channels independently when you don't know what a channel actually represents? I made a node to show ACE-Step 1.0 (audio model) latents visually and it seemed like one of the channels mainly just acted as a squelch function, muting/lowering the volume of the audio when it contained high values. What a channel does could be anything. Obviously audio stuff doesn't apply to your sampler, that's just an example. It would be pretty easy for you to prove me wrong or test how well your classification works. Just VAE encode an image, run your classifier on it and have it output a greyscale image for each channel in the original latent. You also might want to scale it up based on the spatial compression factor (8x for SDXL and SD1.5). Then you can verify if it's classifying areas the way you expect it to, whether the classifications agree per channel, etc. If you're not sure how to do that, it should be pretty easy for a LLM. Actually, I have a node that can do some of the work for you: It would be something like: # Assumes your classify functions, torch, etc are in scope.
class ClassifyMaterialNode:
FUNCTION = "go"
RETURN_TYPES = ("LATENT",)
@classmethod
def INPUT_TYPES(cls) -> dict:
return {
"required": {
"latent": ("LATENT",),
"material_id": ("INT", {"default": 0}),
"normalize": ("BOOLEAN", {"default": False}),
},
}
@classmethod
def go(cls, *, latent: dict, material_id: int, normalize: bool) -> tuple[dict]:
samples = latent["samples"]
responses = _laws_texture_energy(samples, normalize=normalize)
classified = _classify_material(responses)
classification_map = (classified == material_id).to(dtype=samples.dtype)
return ({"samples": classification_map},)Completely untested so you might have to make some changes but that's the general idea. The output latent should have
If you want an example of doing that, you can check https://github.com/blepping/comfyui_dum_samplers/blob/main/py/nodes/simancestral_euler_sampler.py - it builds the ETA map based on cosine similarity with something else but you can use the same approach for whatever. You might also want to look at
If you have a way to do your classification in pixel space then there may be an easier way than training a bunch of classifier models. Use TAESD (or equivalent fast/light previewer) VAE to VAE decode This ended up being a lot longer and involved than I had planned. I hope it's at least somewhat helpful. |
|
Oh boy, a lot to read, thanks for taking the time to write it, really appreciate it! Most people I shared this project with (e.g. on Discord) usually are too lazy to write a decent reply. Indeed, even the last time I pushed that branch, I felt the readme was questionable, but it's just a gut feeling since I have yet to learn how things affect each other, especially for stable diffusion stuff, so your feedback is definitely a good one.
Thanks for considering. Indeed it is also a project I made out of curiosity and I happened to get some free time to do it, so I can relate. If I couldn't get anyone to jump in or send PRs to the repo in the next couple of days (or weeks?), I'm not even sure if I will have the time to check the repo again unless it's the weekend or something like that. So please don't sweat it haha.
No no, it is the kind of reply that I've been waiting for over the past few days actually. These references that you gave are gold to me. I really appreciate that! |
By all means, hope you find them helpful! There's a lot of weird stuff, also some useful tools. Some of it hasn't been updated in some time or was written when I knew a lot less than I do now, so the quality can vary. By the way, I took a look at the abstract for the Laws classification stuff (the full paper didn't seem to be publicly available) and it looks like it's for images. So if you do stuff in pixel space, assuming that works then you have your classifier already. I would recommend doing some testing with it using something like the If you use that approach, you're going to have to deal with loading a VAE and VAE decoding stuff while sampling. I recommend using the TAESD-type models that are used for live previews since they are lighter and faster than the normal VAE. When you load one model, it might result in others getting evicted from memory. In other words, the model you're actually sampling with may get unloaded after you do the VAE decode so you will need to ensure it gets reloaded if necessary before you do the next sampler step. There are some examples of dealing with that in my https://github.com/blepping/comfyui_jankdiffusehigh nodes, look at Not sure if you've considered it, but most of the stuff in your sampler could be implemented as pre or post CFG patch instead since you're essentially just running filters on |
|
@blepping |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Repository: https://github.com/galpt/infinity-diffusion
If you have some free time and want to help test it with your use cases and provide feedback, that would be greatly appreciated.
How to try it
Note
I'd recommend trying the
aetherbranch directly. It shows promising results so far and currently it's set as the default branch.The other branches will stay as-is for a while as a comparison to the
aetherbranch.It installs as a custom node with one command:
Restart ComfyUI and "infinity" appears in both the sampler and scheduler dropdowns. Uninstall is the same command with uninstall instead of install. Nothing inside ComfyUI's own files is touched.
If there is enough interest from the community, I can look into submitting it as a built-in option, or other devs are welcome to add it to ComfyUI or their own projects.
All reactions