Update documentation and examples
This commit is contained in:
blepping
@@ -0,0 +1,316 @@
|
||||
# Advanced Nodes
|
||||
|
||||
## Normalization
|
||||
|
||||
Normalization essentially rebalances the noise (or mixture of noise) to 1.0 strength and then scales based
|
||||
on the factor of the node. Most nodes will allow you to set three values:
|
||||
|
||||
* `default`: By default, noise will be normalized only just before it's used. So you could consider this setting to be false except for where it is connected to an actual noise consumer (i.e. a `SamplerConfigOverride` node).
|
||||
* `forced`: Will always normalize.
|
||||
* `disabled`: Will never normalize.
|
||||
|
||||
## `SONAR_CUSTOM_NOISE`
|
||||
|
||||
This node output type actually constitutes a chain of noise items. For most nodes, when you use it as input,
|
||||
they will add an item to the chain. There are some exceptions that treat the `SONAR_CUSTOM_NOISE` input as a list:
|
||||
|
||||
* `SonarRepeatedNoise`
|
||||
* `SonarRandomNoise`
|
||||
|
||||
There are also some exceptions that will consume the list rather than adding an item to it:
|
||||
|
||||
* `SonarModulatedNoise`
|
||||
* `SonarCompositeNoise`
|
||||
* `SonarScheduledNoise`
|
||||
* `SonarGuidedNoise`
|
||||
|
||||
The distinction is mainly only important when setting `rescale`. Visual example:
|
||||
|
||||
|
||||
|
||||
It may be counter intuitive that there are actually two separate chains here.
|
||||
|
||||
## Examples
|
||||
|
||||
Note on the examples included for some of these nodes:
|
||||
|
||||
The example images included for some of these nodes all have metadata and can be loaded in ComfyUI.
|
||||
Generated using `dpmpp_2s_ancestral`, Karras scheduler and starting out with gaussian noise then switching
|
||||
to the custom noise type at the 35% mark.
|
||||
|
||||
***
|
||||
|
||||
### `SonarCustomNoise`
|
||||
|
||||
You can chain `SonarCustomNoise` nodes together to mix different types of noise. The order of `SonarCustomNoise` nodes is not important.
|
||||
|
||||
Parameters:
|
||||
|
||||
- `factor` controls the strength of the noise.
|
||||
- `rescale` controls rebalancing `factor` for nodes in the chain. When `rescale` is set to `0.0`, no rebalancing will occur. Otherwise the current node as well as the nodes connect to it will have their `factor` adjusted to add up to the rescale value. For example, if you have three nodes with `factor` 1.0 and the last with `rescale` 1.0, then the `factor` value will be adjusted to `1/3 = 0.3333...`. *Note*: Rescaling uses the `factor` absolute value.
|
||||
- `noise_type` allows you to select the built-in noise type.
|
||||
|
||||
***
|
||||
|
||||
### `NoisyLatentLike`
|
||||
|
||||
This node takes a reference latent and generates noise of the same shape. The one required input is `latent`.
|
||||
|
||||
You can connect a `SonarCustomNoise` or `SonerPowerNoise` node to the `custom_noise_opt` input: if that is attached, the built in noise type selector is ignored. The generated noise will be multiplied by the `multiplier` value. **Note**: If you select `brownian` noise (either through the dropdown or by connecting custom noise nodes) you must connect a model and sigmas.
|
||||
|
||||
The node has two main modes: simply generate and scale the noise by the multiplier and return or add it to the input latent. In this mode, you don't connect anything to the `mul_by_sigmas_opt` or `model_opt` inputs and you would use other nodes to calculate the correct strength.
|
||||
|
||||
In the second mode you must connect sigmas (for example from a `BasicScheduler` node) to the `mul_by_sigmas_opt` input and connect a model to the `model_opt` input. It will calculate the strength based on the first item in the list of sigmas (so you could use something like a `SplitSigmas` node to slice them as needed). Note that `multiplier` still applies: the calculated strength will be scaled by it. This second mode is generally this is the most convenient way to use the node since the two main uses cases are: making a latent with initial noise or adding noise to a latent (for img2img type stuff).
|
||||
|
||||
If you want to create noise for initial sampling, connect model and sigmas to the node, connect an empty latent (or one of the appropriate size) to it and that is basically all you need to do (aside from configuring the noise types). For img2img (upscaling, etc), either slice the sigmas at the appropriate or set a denoise in something like the `BasicScheduler` node. *Note*: For img2img, you also need to turn on the `add_to_latent` toggle. Turning this on doesn't matter for initial noise since an empty latent is all zeros.
|
||||
|
||||
**Note**: This node does not currently respect the latent noise mask.
|
||||
|
||||
***
|
||||
|
||||
### `SamplerConfigOverride`
|
||||
|
||||
This node can be used to override configuration settings for other samplers, including the noise type. For example, you could force `euler_ancestral` to use a different noise type. It's also possible to override other settings like `s_noise`, etc. *Note*: The wrapper inspects the sampling function's arguments to see what it supports, so you should connect the sampler directly to this rather than having other nodes (like a different sampler wrapper) in between.
|
||||
|
||||
***
|
||||
|
||||
### `SonarModulatedNoise`
|
||||
|
||||
Experimental noise modulation based on code stolen from
|
||||
[ComfyUI-Extra-Samplers](https://github.com/Clybius/ComfyUI-Extra-Samplers). `intensity` and `frequency` modulation
|
||||
types _probably_ do not work correctly for normal sampling — I expect the modulation will be based on the tensor
|
||||
where the noise sampler was created rather than each step. However it may be useful for something like restart sampling
|
||||
noise (see `KRestartSamplerCustomNoise` below). You can also pass it a reference latent to modulate based on
|
||||
instead (only used for `intensity` and `frequency` modulation types).
|
||||
|
||||
*Note*: It's likely this node will be changed in the future.
|
||||
|
||||
<details>
|
||||
|
||||
<summary>⭐ Expand Example Images ⭐</summary>
|
||||
|
||||
<br/>
|
||||
|
||||
These examples all use the `spectral_signum` modulation type as it doesn't depend on a reference.
|
||||
|
||||
#### Positive Strength
|
||||
|
||||
Dims 3:
|
||||
|
||||
|
||||
|
||||
Dims 3 (with studentt noise):
|
||||
|
||||
|
||||
|
||||
Dims 2:
|
||||
|
||||
|
||||
|
||||
Dims 1:
|
||||
|
||||
|
||||
|
||||
#### Negative Strength
|
||||
|
||||
Dims 3:
|
||||
|
||||
|
||||
|
||||
Dims 3 (with studentt noise):
|
||||
|
||||
|
||||
|
||||
Dims 2:
|
||||
|
||||
|
||||
|
||||
Dims 1:
|
||||
|
||||
|
||||
|
||||
</details>
|
||||
|
||||
***
|
||||
|
||||
### `SonarRepeatedNoise`
|
||||
|
||||
Experimental node to cache noise sampler results. Why would you want to do this? Some noise samplers are
|
||||
relatively slow (`pyramid` for example) or it may be slow to generate noise if you are mixing many types
|
||||
of noise. When `permute` is enabled, a random effect like flipping the noise or rolling it in some dimension
|
||||
will be chosen each time the noise sampler is called. I recommend leaving `permute` on. Note that repeated
|
||||
noise (especially with `permute` disabled) can be stronger than normal noise, so you may need to rescale to
|
||||
a value lower than `1.0` or decrease `s_noise` for the sampler. You may also set the maximum number of
|
||||
times noise is reused by setting `max_recycle`.
|
||||
|
||||
<details>
|
||||
|
||||
<summary>⭐ Expand Example Images ⭐</summary>
|
||||
|
||||
<br/>
|
||||
|
||||
Repeated noise is very strong (especially when permute is disabled). You generally won't get good
|
||||
results using 1.0 strength:
|
||||
|
||||
|
||||
|
||||
I recommend considerably decreasing the strength (example here is using 0.75 which is still a bit too much):
|
||||
|
||||
|
||||
|
||||
</details>
|
||||
|
||||
***
|
||||
|
||||
### `SonarCompositeNoise`
|
||||
|
||||
Allows compositing noise types based on a mask. Noise is mixed based on the strength of the mask at a location.
|
||||
For example, where the mask is 1.0 (max strength) you will get 100% `noise_src` and 0% `noise_dst`. Where the
|
||||
mask is 0.75 you will get 75% `noise_src` and 25% `noise_dst`.
|
||||
|
||||
<details>
|
||||
|
||||
<summary>⭐ Expand Example Images ⭐</summary>
|
||||
|
||||
<br/>
|
||||
|
||||
These examples use a base noise type of gaussian and composite in an area with a different type
|
||||
near middle. The custom noise is also set to a higher strength than normal to highlight the effect.
|
||||
|
||||
**No Composite (for comparison)**
|
||||
|
||||
|
||||
|
||||
**Brownian**
|
||||
|
||||
|
||||
|
||||
**Pyramid**
|
||||
|
||||
|
||||
|
||||
**Pyramid negative factor**
|
||||
|
||||
|
||||
|
||||
</details>
|
||||
|
||||
***
|
||||
|
||||
### `SonarScheduledNoise`
|
||||
|
||||
Allows switching between noise types based on percentage of sampling (note: not percentage of steps).
|
||||
|
||||
**Note**: You don't have to connect the fallback noise type but the default is to generate _no_ noise, which
|
||||
is most likely not what you want. The majority of the time, it is recommend to connect something like gaussian
|
||||
noise at 1.0 strength.
|
||||
|
||||
All the example images here use the `SonarScheduledNoise` node so you can pick any one of them to see it
|
||||
in action!
|
||||
|
||||
***
|
||||
|
||||
### `SonarGuidedNoise`
|
||||
|
||||
Works similarly as described in the [Guidance](../README.md#guidance) section of the main README, however the guidance is applied
|
||||
to the raw noise. You can use `SonarScheduledNoise` to only apply guidance at certain times. Using `euler`
|
||||
mode seems considerably stronger than `linear`. The default value should be reasonable for `euler`, may need to be
|
||||
increased somewhat for `linear`.
|
||||
|
||||
<details>
|
||||
|
||||
<summary>⭐ Expand Example Images ⭐</summary>
|
||||
|
||||
<br/>
|
||||
|
||||
#### Pattern
|
||||
|
||||
These examples use a half circle pattern as the reference: 
|
||||
|
||||
|
||||
##### Euler
|
||||
|
||||
Positive strength:
|
||||
|
||||
|
||||
|
||||
Negative strength:
|
||||
|
||||
|
||||
|
||||
***
|
||||
|
||||
##### Linear
|
||||
|
||||
Normal positive strength:
|
||||
|
||||
|
||||
|
||||
Normal negative strength:
|
||||
|
||||
|
||||
|
||||
Strong positive strength:
|
||||
|
||||
|
||||
|
||||
Strong negative strength:
|
||||
|
||||
|
||||
|
||||
|
||||
***
|
||||
|
||||
#### Gradient
|
||||
|
||||
These examples use a vertical gradient as the reference: 
|
||||
|
||||
That is dark to light. Light to dark examples just flip the gradient vertically.
|
||||
|
||||
##### Euler
|
||||
|
||||
Dark to light:
|
||||
|
||||
|
||||
|
||||
Light to dark:
|
||||
|
||||
|
||||
|
||||
Dark to light (negative strength):
|
||||
|
||||
|
||||
|
||||
Light to dark (negative strength):
|
||||
|
||||
|
||||
|
||||
***
|
||||
|
||||
##### Linear
|
||||
|
||||
Dark to light:
|
||||
|
||||
|
||||
|
||||
Light to dark:
|
||||
|
||||
|
||||
|
||||
Dark to light (negative strength):
|
||||
|
||||
|
||||
|
||||
Light to dark (negative strength):
|
||||
|
||||
|
||||
|
||||
</details>
|
||||
|
||||
***
|
||||
|
||||
### `SonarRandomNoise`
|
||||
|
||||
Randomly chooses between the noise types in the chain connected to it each time the noise sampler is called.
|
||||
You generally do not want to use `rescale` here. You can also set `mix_count` to choose and combine multiple
|
||||
types.
|
||||
@@ -0,0 +1,127 @@
|
||||
# Advanced Power Noise
|
||||
|
||||
## `SonarPowerNoise`
|
||||
|
||||
This node generates [fractional Brownian motion (fBm) noise](https://en.wikipedia.org/wiki/Fractional_Brownian_motion#Frequency-domain_interpretation). It offers versatility in producing various types of noise including gaussian, pink, 2D brownian noise, and all intermediates.
|
||||
|
||||
By default, the node generates normal gaussian noise.
|
||||
|
||||
<details>
|
||||
|
||||
<summary>⭐⭐ Expand advanced parameter explanation ⭐⭐</summary>
|
||||
|
||||
Here's an overview of its parameters:
|
||||
|
||||
- `factor` and `rescale` operate similarly to `SonarCustomNoise`, enabling the addition of multiple sources of noises.
|
||||
- `time_brownian` introduces correlation across sampler timesteps for SDE solvers.
|
||||
- `alpha` is the main parameter. `alpha > 0` amplifies low frequencies; `alpha = 1` yields pink noise, and `alpha = 2` produces brownian noise. Conversely, for `alpha < 0`, it amplifies high frequencies.
|
||||
- `min_freq` and `max_freq` determine the range of frequencies allowed through. Setting `max_freq = `$\sqrt{1/2} \simeq 0.7071$ enables the passage of the highest frequencies. In cases where `alpha < 0`, setting `max_freq = 0.5` is advisable to diminish the power of diagonally oriented frequencies.
|
||||
- `stretch`, `rotate`, and `pnorm` alter the filter's shape by stretching, rotating, or cushioning the band-pass region.
|
||||
- Lowering `mix` moderates the filter's effect by blending back unfiltered gaussian noise from the same sample.
|
||||
- `common_mode` is an attempt to desaturate the latent by injecting the average across channels into every latent channel. **FIXME: it's not all channels anymore** However, this may result in a specific color due to the encoding of the unit vector by the latent space. Note that this is done _after_ the `mix`ing of unfiltered gaussian noise.
|
||||
- `channel_correlation` **FIXME**: TBD
|
||||
- Enabling `preview` provides a visual representation of the filter. `no_mix` sets `mix = 1` for the preview. The preview includes, from left to right:
|
||||
- Fourier domain visualization: Low frequencies at the center, with black indicating filtered-out frequencies.
|
||||
- Spatial visualization of the 2D kernel: The filtering can be interpreted as convolution with the displayed kernel.
|
||||
- Sample: Gaussian sample with shaped frequency spectrum. A single latent channel will look like this.
|
||||
|
||||
**Frequency-domain Interpretation**: The Fourier transform decomposes a 2D latent into sinusoids covering all spatial orientations and frequencies. For an independent and identically distributed gaussian sample, energy is evenly distributed across all frequencies and orientations. Scaling the power spectrum by $1 / f^\alpha$, where $\alpha>0$, boosts low frequencies, introducing spatial correlations.
|
||||
|
||||
**Spatial Domain Interpretation**: A gaussian latent sample comprises independently sampled pixels, exhibiting no spatial correlations. Conversely, a requirement that each pixel value differs from its neighbors by a $\epsilon \sim \mathcal{N}(0, 1)$ results in 2D brownian noise ($\alpha=2$).
|
||||
|
||||
**Seed Considerations**: While the node defaults to outputting gaussian noise, a given seed produce a different sample than the one produced by other gaussian noise sources. This stems from sampling the noise directly in the frequency domain to avoid the cost of a FFT. When `time_brownian = true`, noise sampling occurs in the spatial domain, ensuring that default parameters yield output equivalent to `SonarCustomNoise` set to `brownian`.
|
||||
|
||||
</details>
|
||||
|
||||
<br/>
|
||||
|
||||
From a usage perspective, using positive alpha will tend to create a colorful effect, using negative alpha will create line/streak like artifacts sort of like an oil painting canvas. Start with small values at first (`-0.1`, `0.1`) and adjust as necessary. `time_brownian` makes the effect of power noise (and alpha) stronger - also note that it can only be used when sampling and not for `NoisyLatentLike`. Setting `common_mode` also generally seems to intensify these effects. Different types of models (normal EPS models, v-prediction models, SDXL) generally react differently to these exotic noise types so my advice is to experiment! Lowering `mix` uses normal gaussian noise for part of the generated noise. For example, `mix=1.0` means 100% power noise, `mix=0.5` means 50/50 power noise and normal gaussian noise. This also is about the same as setting factor to `0.5` and plugging in a `SonarCustomNoise` node with factor at `0.5` also and the type set to `guassian`.
|
||||
|
||||
Noise from the `SonarCustomNoise` node and `SonarPowerNoise` can be freely mixed.
|
||||
|
||||
## `SonarPowerFilterNoise`
|
||||
|
||||
This node lets you connect a filter (see below) and a custom noise chain. It basically lets you run any type of noise through the power noise filter.
|
||||
|
||||
New parameters:
|
||||
|
||||
* `filter_norm_factor` controls how much normalization is applied to the filter. `1.0` means fully normalized, `0.0` means no normalization.
|
||||
* You may set the preview type to `custom` to see a color preview of the filtered noise. Note that this uses whatever preview type you have configured in ComfyUI (for example, TAESD). The preview is based on SD 1.5's interpretation of the noise.
|
||||
|
||||
## `SonarPowerFilter`
|
||||
|
||||
Most of the parameters here are similar to the `SonarPowerNoise` node. New parameters:
|
||||
|
||||
* `scale` allows you to scale the filter (you could consider this to be set to `1.0` in the `SonarPowerNoise` node).
|
||||
* `compose_mode` allows you to compose multiple filters. Note that composition occurs like `current_filter OPERATION connected_filter`. So if you set `compose_mode` to `sub`, you will get `current_filter - connected_filter`. Scaling occurs before composition.
|
||||
|
||||
## `SonarPreviewFilter`
|
||||
|
||||
Allows you to preview a filter. It does not modify the input filter.
|
||||
|
||||
***
|
||||
|
||||
## Examples
|
||||
|
||||
The example images are all workflow-included. Generated using `dpmpp_2s_ancestral`, Karras scheduler and
|
||||
starting out with gaussian noise then switching to power noise at the 35% mark. `filter_norm_factor` is set to
|
||||
1.0 in these examples.
|
||||
|
||||
### Node Defaults
|
||||
|
||||
This should be the same as normal gaussian noise.
|
||||
|
||||
|
||||
|
||||
### Positive Alpha
|
||||
|
||||
Positive alpha generally produces a colorful effect. Start with relatively low values and increase
|
||||
until you achieve the desired result. Note that these examples use _relatively_ extreme settings.
|
||||
|
||||
With alpha 0.25:
|
||||
|
||||
|
||||
|
||||
With alpha 0.25, common mode 0.25:
|
||||
|
||||
|
||||
|
||||
With alpha 0.35:
|
||||
|
||||
|
||||
|
||||
With alpha 0.35, common mode 0.35:
|
||||
|
||||
|
||||
|
||||
With alpha 0.5:
|
||||
|
||||
|
||||
|
||||
With alpha 0.5, common mode 0.5:
|
||||
|
||||
|
||||
|
||||
### Negative Alpha
|
||||
|
||||
With alpha -0.5:
|
||||
|
||||
|
||||
|
||||
With alpha -1.5:
|
||||
|
||||
|
||||
|
||||
### Time Brownian Mode
|
||||
|
||||
|
||||
|
||||
With alpha 0.5:
|
||||
|
||||
|
||||
|
||||
With alpha -0.5:
|
||||
|
||||
|
||||
|
||||
|
||||
@@ -0,0 +1,173 @@
|
||||
# Base Noise Examples
|
||||
|
||||
The example images are all workflow-included. Generated using `dpmpp_2s_ancestral`, Karras scheduler and
|
||||
starting out with gaussian noise then switching to the custom noise type at the 35% mark.
|
||||
|
||||
Some of these noise types are too extreme to be used for initial generations or even with pure
|
||||
noise of that type. However you can either schedule the noise type to kick in at a certain percentage
|
||||
(as in these examples) and/or mix it with something a bit more run of the mill. See
|
||||
[advanced_noise_nodes](advanced_noise_nodes.md).
|
||||
|
||||
## Brownian
|
||||
|
||||
This is the default noise type for SDE samplers.
|
||||
|
||||
|
||||
|
||||
***
|
||||
## Gaussian
|
||||
|
||||
This is the default noise type for non-SDE samplers.
|
||||
|
||||
|
||||
|
||||
***
|
||||
|
||||
## Green Test
|
||||
|
||||
This is _probably_ not actually green noise. It produces a very colorful effect, however
|
||||
it's very strong and not really suitable for initial generation.
|
||||
|
||||
|
||||
|
||||
You can also use a negative multiplier to achieve a different effect:
|
||||
|
||||
|
||||
|
||||
***
|
||||
|
||||
## Highres Pyramid
|
||||
|
||||
|
||||
|
||||
Variation using area scaling:
|
||||
|
||||
|
||||
|
||||
Variation using bislerp scaling:
|
||||
|
||||
|
||||
|
||||
***
|
||||
|
||||
## Laplacian
|
||||
|
||||
|
||||
|
||||
***
|
||||
|
||||
## Perlin
|
||||
|
||||
|
||||
|
||||
***
|
||||
|
||||
## Pink
|
||||
|
||||
|
||||
|
||||
***
|
||||
|
||||
## Power Builtin
|
||||
|
||||
|
||||
|
||||
Also see the [Advanced Power Noise](advanced_power_noise.md) examples.
|
||||
|
||||
***
|
||||
|
||||
## Pyramid
|
||||
|
||||
|
||||
|
||||
You can also use a negative multiplier to achieve a different effect:
|
||||
|
||||
|
||||
|
||||
Variation using area scaling:
|
||||
|
||||
|
||||
|
||||
Variation using bislerp scaling:
|
||||
|
||||
|
||||
|
||||
***
|
||||
|
||||
## Pyramid Discount5
|
||||
|
||||
Pyramid noise, generated with a discount of 0.5. (Generally less extreme effect.)
|
||||
|
||||
|
||||
|
||||
***
|
||||
|
||||
## Pyramid Mix
|
||||
|
||||
Pyramid mix is a combination of positive and negative pyramid noise. The effect on
|
||||
the generation is mild compared to raw pyramid noise.
|
||||
|
||||
|
||||
|
||||
You can also use a negative multiplier to achieve a different effect:
|
||||
|
||||
|
||||
|
||||
Variation using area scaling:
|
||||
|
||||
|
||||
|
||||
You can also use a negative multiplier to achieve a different effect:
|
||||
|
||||
|
||||
|
||||
Variation using bislerp scaling:
|
||||
|
||||
|
||||
|
||||
You can also use a negative multiplier to achieve a different effect:
|
||||
|
||||
|
||||
|
||||
***
|
||||
|
||||
## Pyramid Old
|
||||
|
||||
This may not actually be pyramid noise at all. Also note that it is quite slow to generate as it
|
||||
effectively generates noise ~60x the latent size.
|
||||
|
||||
|
||||
|
||||
Variation using area scaling:
|
||||
|
||||
|
||||
|
||||
Variation using bislerp scaling:
|
||||
|
||||
|
||||
|
||||
***
|
||||
|
||||
## Rainbow
|
||||
|
||||
Rainbow is a mix of Perlin and Green noise types.
|
||||
|
||||
The "mild" variation uses a relatively low proportion of green noise:
|
||||
|
||||
|
||||
|
||||
The "intense" variation uses a higher proportion of green noise for a more extreme effect.
|
||||
|
||||
|
||||
|
||||
***
|
||||
|
||||
## Studentt
|
||||
|
||||
|
||||
|
||||
***
|
||||
|
||||
## Uniform
|
||||
|
||||
|
||||
@@ -0,0 +1,40 @@
|
||||
# FreeU Extreme
|
||||
|
||||
I admit it's a really dumb name. This is basically a build-your-own FreeU kit.
|
||||
|
||||
## Example Workflow
|
||||
|
||||
Workflow image is also workflow-embedded.
|
||||
|
||||
|
||||
|
||||
## Nodes
|
||||
|
||||
### `FreeUExtreme`
|
||||
|
||||
Allows you to apply a FreeU (v1 or v2) effect to input, output or middle blocks.
|
||||
|
||||
**Note**: ComfyUI by default does not allow patching the middle in the required way. You will need to have
|
||||
[FreeU Advanced](https://github.com/WASasquatch/FreeU_Advanced) installed and enabled, otherwise connecting
|
||||
configs to the `middle` input will have no effect.
|
||||
|
||||
Also note that input and middle do not have a `skip` target so configs targetting that will never match.
|
||||
|
||||
### `FreeUExtremeConfig`
|
||||
|
||||
Better documentation coming soon hopefully. For now, see the workflow example above to get started.
|
||||
|
||||
Also see documentation on filters [here](./advanced_power_noise.md#sonarpowerfilter).
|
||||
|
||||
## Examples
|
||||
|
||||
ComfyUI built-in FreeU V2 for reference:
|
||||
|
||||
|
||||
|
||||
FreeU Extreme example:
|
||||
|
||||
|
||||
|
||||
Note that this is just for example purposes - no attempt was made to get a pretty generation. You
|
||||
may get better results enabling `hidden_mean` even for the skip connections.
|
||||
Reference in New Issue
Block a user