ComfyUI Flow Matching Inverter Nodes

Custom nodes for ComfyUI that target rectified-flow models (for example, Qwen). The nodes cover three core tasks:

• integrate the model's velocity field to produce a partially inverted latent,
• add noise using the same schedule as the sampler, and
• perform small latent-domain cleanups or conditioning perturbations.

Each node is available on its own, and the hybrid node chains them together for a single-drop workflow.

This node pack also includes a variety of latent and conditioning space noise nodes enabling you to do really interesting and insane things in the latent and conditioning space.

Conditioning primer

ComfyUI passes prompt information around as a Python list of [embedding, metadata] pairs. The embedding tensor stores the per-token hidden states produced by the text encoder (shape ≈ batch × tokens × features, or tokens × features when the batch dimension is implicit). The accompanying metadata dictionary carries auxiliary fields—most notably pooled_output, a single vector that summarises the whole prompt (CLS token, sentence average, etc.).

Whenever you tweak conditioning with the nodes below you are directly manipulating those tensors. Splitting, blurring, or scaling happens along the token axis, while pooled outputs are updated in lockstep so downstream samplers still receive a coherent summary.

Node overview

Qwen Rectified Flow Inverter

Runs the rectified-flow ODE backwards to re-noise an encoded latent. Supports deterministic amplification, seeded stochastic perturbation, optional output normalisation, and shares the same interface as the sampler (steps, conditioning, etc.).

Latent Gaussian Blur

Applies a Gaussian blur directly in latent space. Spatial Only blurs each channel independently; Spatial and Channel performs a joint 3D blur across channels for stronger smoothing.

Latent Frequency Split

Separates a latent into low-pass structure and high-frequency detail bands using a Gaussian crossover so you can process each side differently before recombining.

Add Latent Noise (Seeded)

Adds seeded Gaussian noise scaled by the input latent's standard deviation. Useful whenever a latent needs a controlled bump in noise without touching the schedule.

Latent Perlin Fractal Noise

Generates smooth octave-based Perlin noise and adds it to the latent. Tune the base frequency, octaves, persistence, and lacunarity to decide whether you want broad undulations or fine stippling, and choose between shared or per-channel fields for subtle colour-channel offsets.

Latent Simplex Noise

Produces layered simplex noise—Perlin's successor—for isotropic, artifact-free texture. Frequency, octaves, persistence, and lacunarity work just like the Perlin node, while temporal_mode lets you keep the same pattern across video frames (locked) or reseed each frame (animated).

Latent Worley Noise

Synthesises cellular Worley (Voronoi) patterns by scattering seeded feature points. Adjust the point count, distance metric, and jitter to morph between cracked clay, reptilian scales, or bubbly foams. Combine with animated mode to introduce evolving cellular motion in video latents.

Latent Reaction-Diffusion

Runs a Gray–Scott reaction-diffusion simulation in latent space to generate self-organising Turing patterns. Feed, kill, and diffusion controls expose the classic Gray–Scott parameter space; temporal_mode toggles between a single locked pattern and re-simulating per frame.

Latent Fractal Brownian Motion

Wraps any of the base noise primitives (simplex, Perlin, or Worley) in an fBm stack to build rich multi-scale detail. Combine frequency, octaves, persistence, and lacunarity to decide how aggressively the layers accumulate, and pick Worley for cellular fBm, Perlin for smooth ridges, or simplex for organic grain.

Under the hood the node keeps resampling the chosen base noise at higher frequency while shrinking the amplitude (classic fBm). Each octave is normalized before it is added, so persistence becomes a clean gain control instead of inheriting whatever variance the base noise happens to produce. Worley uses the feature_points count scaled by frequency to keep the cell density consistent as lacunarity climbs. Once all octaves are summed, the composite field is normalized again, multiplied by the latent's standard deviation, and finally scaled by strength before being added back. channel_mode decides whether every channel shares the same Brownian field or gets an independent reseeded stack, and temporal_mode controls whether video batches reuse the field (locked) or generate fresh motion on every frame (animated).

Latent Swirl Noise

Warps the latent around one or more seeded vortices using grid_sample. Choose between a global grid (shared across channels) or per-channel grids, set the fraction of channels to affect, and control vortex count, rotation, falloff radius, centre jitter, and blend to inject painterly whirlpools or gentle twisting motion where you want it.

Conditioning (Add Noise)

Adds seeded Gaussian noise to the CLIP conditioning embeddings (and pooled output when present). Great for introducing gentle prompt variation without rewriting text.

Sensible values of strength range from 0.0 to 1.5. Above this range, the Qwen model's output tends to get silly, with Chinese characters overlaid on the image.

Conditioning (Gaussian Blur)

Smooths the token embeddings with a Gaussian kernel along the prompt sequence, softening sharp emphasis changes while preserving the overall prompt content.

Conditioning (Frequency Split)

Generates low- and high-frequency prompt embeddings so you can tame the overall narrative while separately shaping high-energy emphasis tokens.

Sensible values of sigma range from 0.1 to 1.3. The higher you raise sigma, the more of the conditioning information that will be shoved into the low_pass conditioning output. Values well above 1.0 are really equivalent to knocking out the high pass signal entirely. But strange and interesting results can be obtained by using a very high sigma in combination with a very high value of gain for the high_pass conditioning in the Conditioning (Frequency Merge) node.

Conditioning (Frequency Merge)

Recombines low/high conditioning bands with adjustable gains so you can dial detail back in after sculpting each band independently. Boosting the low band reinforces the prompt's broad narrative and stability, while boosting the high band pushes sharp emphasis changes and punctuation to the forefront.

Feed the output of the Conditioning (Frequency Split) node into the input of this node and then play around with the gain knobs for interesting effects. Or peel off the high or low-pass conditioning and then add noise to one or the other, or combine the high pass conditionings of two different CLIP Text Encode outputs... the combinations are endless.

Conditioning (Scale)

Multiplies the conditioning embeddings and pooled outputs by a user-defined factor so you can mute (0.0), keep (1.0), or amplify (>1.0) prompt influence without editing text.

Forward Diffusion (Add Scheduled Noise)

Uses the model's noise schedule (via KSampler) to add the amount of noise that corresponds to a given progress value. This is the "anchor" latent that samplers expect when starting from a given denoise level.

Latent Hybrid Inverter (Qwen)

Convenience node that calls the inverter and forward diffusion nodes internally, then blends their outputs using spherical interpolation. It keeps separate seeds for the creative and anchor paths and exposes the blend weight so you can bias towards either latent.

Creative play

Image exploration rarely means chasing a single "correct" answer. These nodes are designed to be mixed and layered so you can steer a latent in different, often surprising, directions:

• Lead with Latent Hybrid Inverter to recover structure, then nudge that structure by adding Conditioning (Add Noise) at a subtle strength (≈0.05–0.15). The latent stays coherent while the prompt wanders just enough to spark new ideas.
• Follow the noise with Conditioning (Gaussian Blur) to soften abrupt emphasis changes or to merge multi-prompt blends into a single vibe. Blurring after noise often produces dreamlike, painterly shifts instead of chaotic drift.
• Use Conditioning (Scale) as a volume knob while iterating: dial the factor down to 0.3–0.5 when you want the image to respond mostly to the recovered latent, then crank it above 1.5 when the textual guidance should take the lead.
• Pair Latent Gaussian Blur with Conditioning (Gaussian Blur) for holistic smoothing—latent blur calms texture while conditioning blur calms prompt pacing. Reintroduce energy by sprinkling in Add Latent Noise or ramping the conditioning scale afterwards.
• Sculpt structured detail by layering Latent Perlin Fractal Noise (shared mode for broad waves, per-channel for colour shifts) before blending back with Latent Swirl Noise—increase vortices for multi-spiral motion, switch to per-channel grids, or drop channel_fraction to swirl only a handful of feature maps.
• Split a latent with Latent Frequency Split, sharpen or noise the high band while blurring the low band, then recombine via Latent Mixer/Add nodes to weave sharp detail onto soft composition.
• Carve prompts with Conditioning (Frequency Split) and return them through Conditioning (Frequency Merge): keep low-band text steady for the story while remapping the high band through noise, blur, or scale to emphasize only certain fragments.
• Swap seeds between the inverter, forward diffusion, and noise nodes as you iterate. Small tweaks here can shift the interplay between latent detail and prompt guidance, rewarding playful experimentation.
• Chase naturalistic textures by stacking Latent Simplex Noise or the fBm node before sampling; switch temporal_mode to animated when generating video to keep organic flicker under control. For harder cellular structures, blend in Latent Worley Noise, and when you need life-like stripes or spots, seed Latent Reaction-Diffusion with slightly different feed/kill rates to explore the Gray–Scott pattern atlas.

Treat each slider as a brushstroke: push a setting until it breaks, back off to the sweet spot, and capture the happy accidents along the way.

Example workflows

Manual composition

1. Encode an image to a latent.
2. Feed the latent into Qwen Rectified Flow Inverter (pick strength/amplification/perturbation).
3. Feed the clean latent into Forward Diffusion with the same strength but a different seed.
4. Optionally clean either latent with Latent Gaussian Blur or Add Latent Noise.
5. Blend the creative and anchor latents (e.g. Latent Slerp) and sample with denoise equal to the chosen strength.

Hybrid node quickstart

1. Encode an image to a latent.
2. Run the latent through Latent Hybrid Inverter with your preferred strength and blend factor.
3. Sample the result with denoise matching the passed strength.

Combination example

Check Examples/Combination Example.png for a full workflow that combines the inverter, simplex and Worley noise, reaction-diffusion, and fBm nodes. The diagram illustrates how to stage noise injection around the hybrid inverter and how the new temporal_mode/channel_mode options keep video latents coherent while still layering rich structure.

Installation

1. Clone or download this repository into ComfyUI/custom_nodes/.
2. Restart ComfyUI to load the nodes (look under Qwen/Sampling and Latent).

Running the tests

The repository ships with a helper script that prepares a virtual environment and executes the pytest suite. From the project root run:

./run_tests.sh

The script creates (or reuses) a .venv folder, installs the dependencies from requirements.txt, and then launches pytest. Pass additional arguments to forward them to pytest—for example ./run_tests.sh -k noise runs only the tests whose names include noise.