Nodes/ComfyUI-MaxedOut/Load Latent MXD
ComfyUI Node

Load Latent MXD

Resume a WAN 2.2 run with every setting intact

By Maxed-Out-99·Created about a year ago·Updated 21 days ago· 9
Load Latent MXD
    • shift
    • positive
    • negative
    • samples
    • steps
    • cfg
    • sampler_name
    • scheduler
    • end_at_step
    • filename_prefix
    • trim_latent
    • high_workflow
    latent
    run_folderfalse
    refresh_before_runfalse

    If Save Latent I2V MXD is how you pause a WAN 2.2 generation mid-flight, Load Latent I2V MXD is how you unpause it. It picks a .latent file from your input/latents folder and reconstructs everything the save bundled: the latent itself, the positive and negative conditioning, and the sampler settings - steps, CFG, sampler, scheduler, shift, and end-at-step - so the second stage of your workflow runs with the exact parameters the first stage used. No reconstructing settings from memory, no drift between stages.

    What comes out

    The node reads the .latent safetensors file (and its .cond.pt conditioning sidecar) and returns a stack of outputs:

    • samples - the latent, ready for the sampler
    • positive and negative - the conditioning from the original WAN image→video encode
    • steps, cfg, sampler_name, scheduler, end_at_step - the sampler configuration, restored
    • shift - the SD3-style shift value this workflow uses
    • filename_prefix - the base name derived from the loaded file (folder + stem), so if you save the finished result it lands with a sensible name

    You wire those into a KSampler-Advanced-style node and a save, and you're continuing the run rather than restarting it. The input is just latent - a dropdown of every .latent file in input/latents.

    How it fits the two-stage pattern

    This is the finishing stage of the workflow the save node starts: run stage one, save the high-noise latent, then come back - maybe after a reboot, maybe with a different upscale branch - load it here, and finish the low-noise pass. Because the workflow graph itself is embedded in the file, dragging the finished video back into ComfyUI can reconstruct the whole two-stage setup on one canvas. The pair is built for the "save your place in an expensive generation" problem that long WAN 2.2 runs make painfully real.

    Gotchas

    • Keep .latent and .cond.pt together in the same folder. The loader needs the sidecar for conditioning; split them and you lose the settings that make the resume seamless.
    • Files must live under ComfyUI/input/latents - that's where the dropdown scans, recursively by folder.
    • It's built for WAN 2.2's latent format and the pack's save convention. It'll load other .latent files, but the settings reconstruction is tuned for this pack's pipeline.

    Install

    ComfyUI Manager → search Maxed Out → install → restart, or:

    cd ComfyUI/custom_nodes
    git clone https://github.com/Maxed-Out-99/ComfyUI-MaxedOut
    

    No requirements.txt, no model downloads. Under MXD/Latents (I2V). If you've ever re-run a long generation from scratch just because you didn't want to redo the setup, this is the node that makes "resume later" a real option.

    CategoryMXD/Latents

    Inputs (3)

    NameTypeDefaultDescription
    latentCOMBO0 options:
    run_folderBOOLEANfalseWhen enabled, hitting Queue Prompt auto-queues every latent in this file's folder, one after another, instead of just the selected file.
    refresh_before_runBOOLEANfalseWhen run_folder is on, re-scan the latents folder for new files right before the queuing loop starts, instead of using the dropdown list as of whenever it was last populated. Use this when another workflow is still writing latents into this folder as you queue this one.

    Outputs (12)

    NameTypeDescription
    shiftFLOAT
    positiveCONDITIONING
    negativeCONDITIONING
    samplesLATENT
    stepsINT
    cfgFLOAT
    sampler_nameeuler,euler_cfg_pp,euler_ancestral,euler_ancestral_cfg_pp,heun,heunpp2,exp_heun_2_x0,exp_heun_2_x0_sde,dpm_2,dpm_2_ancestral,lms,dpm_fast,dpm_adaptive,dpmpp_2s_ancestral,dpmpp_2s_ancestral_cfg_pp,dpmpp_sde,dpmpp_sde_gpu,dpmpp_2m,dpmpp_2m_cfg_pp,dpmpp_2m_sde,dpmpp_2m_sde_gpu,dpmpp_2m_sde_heun,dpmpp_2m_sde_heun_gpu,dpmpp_3m_sde,dpmpp_3m_sde_gpu,ddpm,lcm,ipndm,ipndm_v,deis,res_multistep,res_multistep_cfg_pp,res_multistep_ancestral,res_multistep_ancestral_cfg_pp,gradient_estimation,gradient_estimation_cfg_pp,er_sde,seeds_2,seeds_3,sa_solver,sa_solver_pece,ddim,uni_pc,uni_pc_bh2
    schedulersimple,sgm_uniform,karras,exponential,ddim_uniform,beta,normal,linear_quadratic,kl_optimal
    end_at_stepINT
    filename_prefixSTRING
    trim_latentINT
    high_workflowSTRING