MiniMax H3 Add Noise (resample):
MiniMax H3 Add Noise (resample) prepares an existing latent for a second sampling pass with a CONST/flow-parameterized model such as MiniMax H3. It adds noise using the model’s own scaling behavior, then applies an inverse scaling step when the model provides one. This compensates for scaling that the second sampler performs even when its noise source is disabled. The goal is to let the second pass refine the result, including at higher denoise settings, instead of turning it into an almost entirely noisy result.
MiniMax H3 Add Noise (resample) Input Parameters:
model
The model supplies the sampling behavior and latent conversions used to prepare the result. Use the model intended for the second sampling pass so that the noise preparation matches that pass. This is a required MODEL connection; the node defines no default, minimum, maximum, or selectable options.
noise
The NOISE connection supplies the noise that the node generates and adds to the latent. It determines the noise contribution used to restart sampling, while the model and sigma schedule determine how that contribution is scaled. This is a required connection; the node defines no default, minimum, maximum, or selectable options.
sigmas
The SIGMAS connection supplies the sampling schedule from which the node calculates its noise scale. With more than one value, the scale is the absolute difference between the first and last values; with one value, that value is used. An empty schedule makes the node return the supplied latent unchanged. Use the schedule intended for the continuation pass. This is a required connection; the node defines no default, minimum, maximum, or selectable options.
latent_image
The LATENT connection supplies the existing latent samples to prepare for continued sampling. When its samples contain nonzero values, the node first converts them into the model’s input latent representation; an all-zero latent skips that conversion. The node preserves other entries in the latent container and replaces its samples with the prepared result. This is a required connection; the node defines no default, minimum, maximum, or selectable options.
MiniMax H3 Add Noise (resample) Output Parameters:
LATENT
The single LATENT output contains the prepared samples for the continuation pass. For a nonempty sigma schedule, these samples have been combined with generated noise, processed through the model’s scaling behavior, converted back to its output latent representation, and sanitized so NaN or infinite values become zero. Other entries from the supplied latent are retained. If the sigma schedule is empty, the output is the supplied latent unchanged.
MiniMax H3 Add Noise (resample) Usage Tips:
- Place this node between the first pass’s latent result and a second
SamplerCustomAdvancedpass that usesDisableNoise. The node prepares the noise contribution; the second sampler should not generate another one. - Connect the model and sigma schedule intended for that continuation pass. Changing the schedule changes the scale calculated by this node and therefore the prepared latent.
- For a packed video-and-audio workflow with already separated branches, use a separate instance for each branch with that branch’s own latent and pre-shifted sigmas, as described by the node’s source comment.
- Check that
sigmasis not empty when you expect the node to add noise. An empty schedule produces an unchanged latent rather than an error.
MiniMax H3 Add Noise (resample) Common Errors and Solutions:
Missing required input: model, noise, sigmas, or latent_image
- Explanation: All four sockets are required. The exact validation message may vary by ComfyUI version.
- Solution: Connect a compatible MODEL, NOISE, SIGMAS, and LATENT source to the corresponding sockets.
KeyError: 'samples'
- Explanation: The supplied latent does not contain the
samplesentry that this node reads. - Solution: Connect a valid LATENT output rather than a plain tensor or another data type.
AttributeError: object has no attribute 'generate_noise'
- Explanation: The
noiseconnection does not provide the noise-generation method the node calls. - Solution: Use a compatible NOISE output and check that the connection has not been replaced by an incompatible custom node.
AttributeError: object has no attribute 'get_model_object'
- Explanation: The
modelconnection does not provide the model interface this node expects. - Solution: Connect a compatible MODEL output and check that it is the model intended for the continuation pass.
Shape mismatch or tensor size mismatch during noise_scaling
- Explanation: The generated noise, latent samples, or model sampling configuration may have incompatible shapes. The exact error text depends on the underlying operation.
- Solution: Check that the latent and noise belong to the same branch and that the model supports that branch’s latent shape. In a separated video-and-audio workflow, use one node per branch with its corresponding sigmas.
The output appears unchanged
- Explanation: When
sigmasis empty, the node explicitly returnslatent_imagewithout processing it. An unchanged-looking result alone does not establish an error. - Solution: Check that a nonempty SIGMAS schedule reaches the node and that the continuation pass is connected to its LATENT output.
The second pass produces an almost entirely noisy result
- Explanation: The continuation setup may not match the workflow this node is designed for, or the model may not provide the inverse scaling method that performs the stated compensation. The source does not establish a single diagnostic error message for this symptom.
- Solution: Confirm that the second
SamplerCustomAdvancedpass receives this node’s output, usesDisableNoise, and uses the intended compatible model and sigma schedule.
