Anima Ref Guider:
AnimaRefGuider, displayed as “Anima Ref Guider” in the Anima Ref Adapter category, controls how text instructions and image references guide image generation in an Anima reference-adapter workflow. It lets you balance prompt direction with reference influence, including character and style references when using a compatible two-reference adapter. Rather than generating an image itself, it prepares guidance for the sampling process—the sequence of steps that gradually turns noise into an image. Its approach builds on classifier-free guidance, which compares predictions made with positive and negative conditioning, and adds adjustable reference guidance to strengthen reference influence. You can also limit reference guidance to part of the generation process and select a correction method intended to keep colors under control at stronger guidance settings. This makes it useful when you want recognizable reference characteristics while retaining creative control through your prompt and generation settings.
Anima Ref Guider Input Parameters:
model
A required MODEL connection supplying the model that the guider will use during sampling. The node creates its guidance controller around this model; it does not load an adapter or encode reference images itself. In an Anima reference-adapter workflow, connect the model prepared for the compatible adapter you intend to use, so the sampling process can interpret the reference information correctly. Model and adapter compatibility matters because supplying reference latents alone does not establish the required model-side support. This connection has no numeric range or default value.
positive
A required CONDITIONING connection containing the positive text conditioning that describes what you want to generate. The guider combines this with the supplied reference information, allowing your prompt to direct the scene while the reference contributes visual characteristics. For example, your prompt can describe a character’s pose, clothing, or surroundings while the reference provides character or style information. Its influence is controlled through text guidance and balanced against reference guidance; stronger text guidance does not guarantee that every prompt detail will override a reference. This connection has no numeric range or default value.
negative
A required CONDITIONING connection containing the negative text conditioning used by classifier-free guidance. In a typical workflow, this represents unwanted characteristics or other negative-prompt instructions and provides the comparison branch used to calculate text guidance. It is part of the guider’s reference-aware conditioning setup, so it is not simply an independent list of filters applied after generation. Connect valid negative conditioning even when your workflow uses an empty negative prompt. This connection has no numeric range or default value.
cfg
A required FLOAT controlling text guidance, with a default of 4.5, minimum of 0.0, maximum of 20.0, and step size of 0.1. This determines the strength of classifier-free text guidance, helping control how strongly the sampling process follows the difference between positive and negative conditioning. Increasing it generally emphasizes prompt direction, while decreasing it reduces that guidance pressure. It is separate from reference guidance, so increasing this value is not the same as increasing reference fidelity. Strong text and reference guidance together can create excessive color intensity or other visual problems, which is why the node also offers correction methods.
ref_guidance
A required FLOAT controlling reference guidance, with a default of 4.5, minimum of 0.0, maximum of 20.0, and step size of 0.1. A value of 1.0 places the reference in both classifier-free guidance branches without applying additional reference amplification; it does not mean that the reference is disabled. Values above 1.0 push the result more strongly toward the reference. The supplied context explains that this extra guidance requires an additional model evaluation per active sampling step compared with keeping the reference in both branches without extra guidance. The allowed range includes values below 1.0, but the provided interface does not fully document their visual behavior, so 1.0 is the clearest baseline for testing reference influence without extra amplification.
correction
A required selection with the options apg, rescale, and none, listed in that order. This selects the correction approach used to help keep colors under control when text guidance and reference guidance are strong. The apg and rescale choices provide different correction methods, while none selects no correction. The supplied node class does not expose the correction formulas, so the precise visual differences should be judged through controlled comparisons rather than assumed from the option names alone. No explicit default is declared in the input metadata; apg is the first listed option.
rescale_phi
A required FLOAT supplied to the guider’s correction settings, with a default of 0.5, minimum of 0.0, maximum of 1.0, and step size of 0.05. This parameter is associated with rescale correction and provides control over that correction setting independently of text guidance and reference guidance. It should not be treated as another reference-strength slider. The supplied class passes this value to the guider but does not show the underlying correction implementation, so it does not establish the exact effect of each endpoint or whether this value affects correction modes other than rescale. Start with 0.5 and compare nearby values when evaluating rescale correction.
ref_latent
An optional LATENT connection containing the reference representation used by a single-reference adapter, or the character reference used by a two-reference adapter. A latent is an encoded representation of an image rather than the image pixels themselves, so this socket needs reference data prepared by an appropriate latent-producing workflow, such as the companion Anima Ref Latent node. This connection supplies reference information; its influence is then governed by the guidance and timing settings. Its default is no connection, represented internally as None, and it has no numeric range. Although individually optional, at least one of ref_latent or style_latent must be connected.
style_latent
An optional LATENT connection supplying the style reference for a two-reference adapter. It allows a compatible adapter to receive style information separately from the character reference, which is useful when you want character characteristics from one image and visual styling from another. The interface explicitly supports connecting the same latent to both reference sockets when you want character and style information from the same image. Its default is no connection, represented internally as None, and it has no numeric range. At least one reference socket must be connected, and accepting this input at the guider level does not establish that every adapter supports a separate style reference.
ref_start
An optional FLOAT defining the beginning of the reference-guidance window as a fraction of the denoising process, with a default of 0.0, minimum of 0.0, maximum of 1.0, and step size of 0.01. A value of 0.0 starts reference guidance at the first sampling step; larger values delay its start until later in generation. This is a relative position in the denoising schedule, not a duration in seconds or an absolute step number. Before this point, the behavior follows the outside setting, so delaying reference guidance does not necessarily remove all reference influence. This value must be less than or equal to ref_end.
ref_end
An optional FLOAT defining the end of the reference-guidance window as a fraction of the denoising process, with a default of 1.0, minimum of 0.0, maximum of 1.0, and step size of 0.01. A value of 1.0 keeps reference guidance active through the last step, while a lower value ends it earlier and hands the remaining steps to the selected outside-window behavior. The node’s tooltip notes that identity and style are established early and suggests ending around 0.6 to save one model evaluation per remaining step when extra reference guidance would otherwise be active. This is a practical starting point rather than a guarantee of identical results, and this value must be greater than or equal to ref_start.
outside
An optional selection controlling sampling steps outside the interval from ref_start to ref_end. The options are reference, no extra guidance and no reference, with reference, no extra guidance used as the method’s default. The first option retains the reference in both classifier-free guidance branches but stops applying extra reference guidance; the tooltip describes this as two model calls per step. The second option drops the reference outside the window and uses plain text classifier-free guidance. Choose the first when you want reference conditioning to remain present after extra guidance ends, or the second when you deliberately want reference-free sampling outside the chosen interval.
Anima Ref Guider Output Parameters:
guider
A GUIDER object containing the model, positive and negative conditioning, reference conditioning, and guidance settings prepared by this node. Connect it to a sampling node that accepts a GUIDER input, such as a compatible advanced sampling workflow. It tells the sampler how to combine text and reference influence during denoising, including when extra reference guidance is active and what behavior to use outside that interval. It is not a generated image, a reference latent, or a modified model output; it is the guidance controller used during subsequent sampling. The node returns exactly one such object after checking that a reference is provided and that the timing interval is valid.
Anima Ref Guider Usage Tips:
- Start with
cfg = 4.5,ref_guidance = 4.5,rescale_phi = 0.5, and the full window fromref_start = 0.0toref_end = 1.0. Keep your seed and other sampling settings fixed while changing one guidance setting at a time, so you can identify what actually improves the result. - Test
ref_guidance = 1.0as a baseline. The reference remains active, but extra reference amplification is removed; this helps you judge whether stronger guidance is needed and avoids the additional evaluation associated with extra reference guidance. - For a single-reference adapter, supply the reference through
ref_latent. For a compatible two-reference adapter, useref_latentfor the character reference andstyle_latentfor the style reference; connect the same latent to both when you want both kinds of information from one image. - Try
ref_start = 0.0,ref_end = 0.6, andoutside = reference, no extra guidanceto reduce the cost of extra reference guidance in later steps while retaining reference conditioning. Compare the result against a full-window run before adopting it for final renders. - Choose
outside = no referenceonly when you want the reference removed outside the active window. Ending extra guidance and removing the reference entirely are different operations, and this option explicitly selects the latter. - If colors become too intense at strong guidance settings, compare
apgandrescalecorrection with the same seed. Also try reducingcfgorref_guidancerather than relying entirely on correction to compensate for excessive guidance. - Use a sampler with a GUIDER socket. This node does not replace the image-encoding, noise, sampling-schedule, or image-decoding parts of your workflow.
Anima Ref Guider Common Errors and Solutions:
Anima Ref Guider needs ref_latent, style_latent, or both
- Explanation: This exact error is raised when neither reference input is connected. Both sockets are individually optional because different reference configurations are possible, but the node requires at least one reference.
- Solution: Connect a valid reference latent to
ref_latent,style_latent, or both, according to your adapter’s supported reference arrangement. For a single-reference adapter, useref_latent.
Anima Ref Guider: ref_start (...) is after ref_end (...)
- Explanation: This error is raised when
ref_startis greater thanref_end. The actual message inserts your numeric values into the parentheses, indicating that the reference-guidance window ends before it starts. - Solution: Set
ref_startto a value less than or equal toref_end. For example, use 0.0 and 1.0 for the full process, or 0.0 and 0.6 to end extra reference guidance earlier.
Required input missing or incompatible connection type
- Explanation: This is a workflow-validation problem rather than an exact error string defined by this node. It can occur when a required input is absent or a socket receives the wrong data type, such as an IMAGE instead of a LATENT or conditioning in place of a MODEL.
- Solution: Check all required connections and settings. Supply MODEL data to
model, CONDITIONING data topositiveandnegative, and encoded LATENT data to the reference sockets.
Model, adapter, or reference compatibility failure during sampling
- Explanation: The provided node class does not define a specific error message for compatibility failures. Problems may appear later when the sampler evaluates the model, because this node prepares guidance but does not itself load the adapter or validate every model-and-reference combination.
- Solution: Confirm that the model has been prepared with the intended compatible Anima reference adapter and that the reference latents were prepared for that workflow. Match single-reference or two-reference connections to the adapter you loaded, and inspect the sampler’s actual error message if the problem persists.
GUIDER output cannot connect to the selected sampler input
- Explanation: This is a socket-type mismatch rather than an exception explicitly raised by the node. The output is a GUIDER object, so it cannot be used where a sampler expects a MODEL, IMAGE, or LATENT.
- Solution: Use a sampling node with a GUIDER input and connect
guiderthere. Keep the model, starting latent, noise, and other sampling components connected to their appropriate sockets in the surrounding workflow.
