JK Easy HiRes Fix:
EasyHRFix performs a high-resolution refinement pass on an existing latent image. It decodes the latent into an image, enlarges that image with an installed pixel-upscaling model, adjusts it to the requested scale, and encodes it back into a latent. It then runs an image-to-image sampling pass to refine the enlarged result. This approach gives you a way to increase image size while letting the model add or revise detail, rather than relying on enlargement alone. You can also add noise before enlargement and supply extra text for the refinement pass. The node returns a latent, so you can continue processing it or decode it elsewhere in your workflow.
JK Easy HiRes Fix Input Parameters:
model
The required MODEL connection supplies the diffusion model used for the final sampling pass. It is also passed to the optional extra-text encoder when you provide extra positive or negative text. Choose a model that works with the conditioning and VAE in your workflow. No default, minimum, or maximum is defined.
clip
The required CLIP connection supplies the text encoder used when either optional extra-text field is nonempty. If both fields are empty, the supplied positive and negative conditioning are used without this extra encoding step. No default, minimum, or maximum is defined.
vae
The required VAE connection decodes the incoming latent into an image and encodes the enlarged image back into a latent for sampling. Using the appropriate VAE for your model and workflow matters because both conversions affect the image passed to refinement. No default, minimum, or maximum is defined.
seed
The random seed used by the refinement sampler and, when applicable, the extra-noise generator and extra-text encoder. Reusing a seed helps you compare setting changes, although changing other settings can still change the result. Default: 0; minimum: 0; maximum: 18446744073709551615.
steps
The number of steps configured for the final denoising pass. More steps give the sampler more iterations to refine the enlarged latent, but generally take more processing time; the visible effect also depends on the denoise setting and sampler. Default: 20; minimum: 1; maximum: 10000.
cfg
The classifier-free guidance scale for the final sampling pass. It controls how strongly sampling follows the supplied conditioning: higher values generally emphasize it more, while excessively high values can harm image quality. Default: 8.0; minimum: 0.0; maximum: 100.0; interface step: 0.1.
sampler_name
Selects the sampling algorithm for the final refinement pass. The available choices come from comfy.samplers.KSampler.SAMPLERS in your ComfyUI installation, so they are not a fixed list in this node. Different algorithms can produce different visual results and processing times. No default, minimum, or maximum is specified in the provided interface.
scheduler
Selects how the final sampling pass schedules its denoising steps. The available choices come from comfy.samplers.KSampler.SCHEDULERS in your ComfyUI installation and can affect the character of the refinement. No default, minimum, or maximum is specified in the provided interface.
positive
The required CONDITIONING connection describing what you want the refinement pass to include. It is used directly when extra_positive_text is empty; otherwise, extra_positive_conditioning_mode determines how the newly encoded text relates to it. No default, minimum, or maximum is defined.
negative
The required CONDITIONING connection describing what you want the refinement pass to avoid. It is used directly when extra_negative_text is empty; otherwise, extra_negative_conditioning_mode determines how the newly encoded text relates to it. No default, minimum, or maximum is defined.
upscale_by
Sets the target enlargement factor relative to the image decoded from the incoming latent. The node first applies the selected pixel-upscaling model, then resizes that result when the model’s native scale differs from this target. A larger factor produces a larger image for the later encoding and sampling stages and can require more memory. Default: 1.5; minimum: 1.05; maximum: 4.0; interface step: 0.05.
latent_image
The required LATENT connection containing the image to enlarge and refine. The node uses its sample dimensions when generating optional extra noise, then decodes it for pixel upscaling. Supply a latent from an earlier generation or processing stage; no default, minimum, or maximum is defined.
denoise
Controls how strongly the final image-to-image sampling pass changes the enlarged latent. Lower values generally retain more of the enlarged image, while higher values allow more substantial revision during refinement. Default: 0.5; minimum: 0.0; maximum: 1.0; interface step: 0.01.
latent_upscaler
Selects the resize method used after pixel-model upscaling when the model’s native scale does not match upscale_by. Despite this parameter’s name, the implementation passes the selected method to ImageScaleBy for the decoded image before VAE encoding. Choices consist of the methods exposed by LatentUpscaleBy.INPUT_TYPES()["required"]["upscale_method"][0], plus lanczos; their exact installed list is resolved at runtime. The declared default is lanczos. This setting has no effect when no adjustment after pixel-model upscaling is needed.
pixel_upscaler
Selects the installed upscale model used to enlarge the decoded image. Its available filenames are obtained at runtime from folder_paths.get_filename_list("upscale_models"); the provided source does not establish a universal set of choices or a default. The chosen model’s native scale also determines whether an additional resize is needed to reach upscale_by.
noise_mode
Selects the noise mode passed to the final Inspire sampler and, when extra noise is injected, to the noise generator. Options: GPU(=A1111) and CPU. Default: GPU(=A1111). This setting is not itself an on/off switch for extra-noise injection.
inject_extra_noise
Controls whether the node attempts to add noise to the incoming latent before decoding and upscaling. Options: disable and enable; default: enable. Injection occurs only when this is enable and extra_noise_strength is greater than zero, and it requires the specified BNK nodes to be installed.
extra_noise_strength
Sets the strength passed to the optional noise-injection step. Higher values apply stronger injected noise before the image is decoded; the final visual effect depends on the subsequent upscaling and sampling. A value of 0.0 skips injection even when inject_extra_noise is enabled. Default: 2.5; minimum: 0.0; maximum: 100; interface step: 0.1.
extra_positive_text
Optional multiline text to encode as additional positive conditioning for the refinement pass. When it is nonempty, the node uses ImpactWildcardEncode and applies extra_positive_conditioning_mode to the resulting conditioning and the supplied positive connection. When empty, this extra encoding step is skipped. Default: an empty string; dynamic prompts are disabled for this field in the interface.
extra_positive_conditioning_mode
Determines how nonempty extra_positive_text affects positive conditioning. Options: replace, which uses the newly encoded conditioning; combine, which combines the supplied and new conditioning; concat, which concatenates them; and average, which calls ConditioningAverage with a weight of 0.5. Default: replace. This setting has no effect while extra_positive_text is empty.
extra_negative_text
Optional multiline text to encode as additional negative conditioning for the refinement pass. When it is nonempty, the node uses ImpactWildcardEncode and applies extra_negative_conditioning_mode to the resulting conditioning and the supplied negative connection. When empty, this extra encoding step is skipped. Default: an empty string; dynamic prompts are disabled for this field in the interface.
extra_negative_conditioning_mode
Determines how nonempty extra_negative_text affects negative conditioning. Options: replace, which uses the newly encoded conditioning; combine, which combines the supplied and new conditioning; concat, which concatenates them; and average, which calls ConditioningAverage with a weight of 0.5. Default: replace. This setting has no effect while extra_negative_text is empty.
JK Easy HiRes Fix Output Parameters:
LATENT
The refined, enlarged latent produced by the final Inspire sampling pass. It represents the result after optional noise injection, VAE decoding, pixel-model upscaling, any resize needed to meet the target scale, VAE encoding, and image-to-image refinement. It is not a directly viewable image: connect it to a compatible downstream latent-processing node or VAE decoder to inspect or use the result.
JK Easy HiRes Fix Usage Tips:
- Start with the declared
upscale_bydefault of1.5anddenoisedefault of0.5, then change one setting at a time. Increasedenoisewhen you want the sampling pass to revise the enlarged image more; decrease it when preserving the starting image matters more. - Select an installed
pixel_upscalerbefore running the node. The model performs the first enlargement, whileupscale_bysets the final target scale; a resize occurs only if those scales differ. - To evaluate upscaling without the optional pre-upscale noise step, set
inject_extra_noisetodisableorextra_noise_strengthto0.0. This also avoids the BNK-node requirement for that step. - Leave both extra-text fields empty when your existing positive and negative conditioning already express the refinement you want. Add text only when you need to change that conditioning;
replace, the default mode, substitutes the newly encoded conditioning rather than adding it to the existing connection. - Keep the seed fixed while comparing samplers, schedulers, conditioning modes, or denoise values. This makes it easier to attribute visible differences to the setting you changed, though identical results are not guaranteed across different configurations.
- Higher target scales increase the size of the image handled by later stages. If a run exceeds available memory, try a smaller
upscale_byvalue or a smaller incoming latent.
JK Easy HiRes Fix Common Errors and Solutions:
[ERROR] You need to install 'ComfyUI-Inspire-Pack'
- Explanation: The node checks for
KSampler //Inspireat the start of every run. Without it, the final refinement pass cannot proceed. - Solution: Install or restore ComfyUI-Inspire-Pack, restart ComfyUI so its nodes load, and confirm that
KSampler //Inspireis available.
[ERROR] You need to install 'ComfyUI-Impact-Pack'
- Explanation: At least one extra-text field is nonempty, but
ImpactWildcardEncodeis unavailable. That node is required to encode the extra text. - Solution: Install or restore ComfyUI-Impact-Pack and restart ComfyUI, or leave both extra-text fields empty if you do not need them.
'BNK_NoisyLatentImage', 'BNK_InjectNoise' nodes are not installed.
- Explanation:
inject_extra_noiseis enabled andextra_noise_strengthis above zero, but one or both nodes required for injection are unavailable. - Solution: Install the extension that provides both named BNK nodes and restart ComfyUI. To run without this step, set
inject_extra_noisetodisableorextra_noise_strengthto0.0.
[ERROR] Unrecognized extra_positive_conditioning_mode "..."
- Explanation: Nonempty extra positive text was supplied with a mode other than
replace,combine,concat, oraverage. This error is defined in the source, although the standard dropdown presents the supported modes. - Solution: Set
extra_positive_conditioning_modeto one of the four listed options; check any workflow data that sets the value outside the dropdown.
[ERROR] Unrecognized extra_negative_conditioning_mode "..."
- Explanation: Nonempty extra negative text was supplied with a mode other than
replace,combine,concat, oraverage. This error is defined in the source, although the standard dropdown presents the supported modes. - Solution: Set
extra_negative_conditioning_modeto one of the four listed options; check any workflow data that sets the value outside the dropdown.
Upscale model unavailable or failed to load
- Explanation: The selected
pixel_upscalerfilename may no longer be present, orUpscaleModelLoadermay be unable to load that model. The exact error text comes from the loader and is not specified by EasyHRFix. - Solution: Choose a model currently shown in the
pixel_upscalerdropdown. If the expected model is missing, verify that it is installed in ComfyUI’s upscale-model location and refresh or restart ComfyUI.
Out of memory during upscaling or refinement
- Explanation: The enlarged image and latent can demand more memory than the available hardware provides. The exact message depends on the failing operation and is not defined by this node.
- Solution: Reduce
upscale_byor feed the node a smaller latent, then try the run again.
