Skip to content

Ideogram 4#

Ideogram 4 is a 9.3-billion-parameter DiT, and text rendering is the entire reason it is in the catalog. If you need readable words in the picture -- a shop sign, a book cover, a logo, a jersey number, a price label -- this is the family to use. For anything without text in it, most other families are faster and give you more control.

Four versions are available. Three are presets over the same weights; the fourth is a separate distilled checkpoint.

At a glance#

Version Steps CFG Polish steps Sampler Scheduler Base res Batch Negative Detailer USDU
Ideogram 4 (default preset) 20 3 2 euler simple 1216x1216 2 No Refused Refused
Ideogram 4 (Fast) 20 1 none euler simple 1024x1024 2 No Refused Refused
Ideogram 4 (Quality) 48 3 3 euler simple 2048x2048 1 No Refused Refused
Ideogram 4 (Turbo) 12 3 1 euler simple 1024x1024 2 No Refused Refused

Recommended ranges on the three presets are 12 to 48 steps and CFG 1 to 7. All four appear in the picker under one Ideogram family option that expands to Ideogram 4 -> Fast -> Quality -> Turbo. None of them pins a default upscaler, so hi-res fix falls back to the global 4x-AnimeSharp.pth. All four carry a non-commercial licence.

Turbo, Ideogram 4 and Quality are the same weights with different sampler schedules -- switching between them costs nothing but time. Each also runs a short "polish" stage at the end (1, 2 and 3 steps respectively) at a lower guidance scale.

Fast is a genuinely different, distilled checkpoint. It makes one pass through the transformer per step instead of two, which is why it is much quicker than the base model at the same step count, and it runs at CFG 1 by design.

This family is heavy

The base Ideogram models load two transformers at once (a conditional and an unconditional branch) for their asymmetric guidance. The ~8 GB text encoder is pushed onto the CPU so it does not compete for VRAM. Ideogram jobs also lift the GPU power cap for the duration of the render, so expect a hotter, louder card than usual, and expect the queue to move slowly while one is running.

The models#

Ideogram 4#

The default preset: 20 steps with a 2-step polish, at a 1216x1216 base. The middle setting, and the one to start on.
Architecture Ideogram 4, 9.3B single-stream DiT (fp8), dual-model asymmetric CFG
Subtype none
Prompting Prose, short and concrete -- the enhancer rewrites it either way
Negatives Ignored -- the negative branch is a zeroed copy of the positive
Steps 20 (12-48)
CFG 3 (1-7)
Sampler euler
Scheduler simple
Base resolution 1216x1216
Default batch 2
Default upscaler None pinned -- falls back to 4x-AnimeSharp.pth

Good at: short quoted strings rendered legibly (signs, labels, covers, numbers), several distinct pieces of text in one image, photoreal scenes with a deliberate layout, typography that sits correctly on a surface instead of floating.

Weak at: giving you control -- prompt enhancement is forced on, so the caption the model sees is not the sentence you typed; anything you want excluded, since the negative prompt is discarded; fixing faces afterwards, because the Detailer refuses this architecture outright; and anime or illustration work, which the enhancer will quietly steer back towards photography unless you name a medium.

Ideogram 4 (Fast)#

A separate distilled checkpoint (INT8 quantized) that folds guidance into a single branch. Same 20 steps, roughly half the work per step.
Architecture Ideogram 4 Fast, QAD-distilled single-branch DiT (INT8 ConvRot)
Subtype fast, distilled
Prompting Prose, short and concrete -- the enhancer rewrites it either way
Negatives Ignored
Steps 20 (15-30)
CFG 1 (locked -- the recommended range is 1 to 1)
Sampler euler
Scheduler simple
Base resolution 1024x1024
Default batch 2
Default upscaler None pinned -- falls back to 4x-AnimeSharp.pth

Good at: keeping the family's text rendering while running much faster than the base weights, loading lighter (one transformer instead of two), drafting a layout before committing to Quality.

Weak at: guidance -- CFG is pinned at 1 and its recommended range is literally 1 to 1, so raising cfg: buys you nothing; small type and dense text, where the base weights hold up better; LoRA strength balance, since with only one branch the LoRAs stack on the conditional side alone; and it starts at 1024 rather than 1216, so you are also rendering fewer pixels.

Ideogram 4 (Quality)#

48 steps with a 3-step polish at a 2048x2048 base. The slowest job on the whole wiki and, near square ratios, the sharpest text.
Architecture Ideogram 4, 9.3B single-stream DiT (fp8) -- same weights as the default preset
Subtype none
Prompting Prose, short and concrete -- the enhancer rewrites it either way
Negatives Ignored
Steps 48 (12-48)
CFG 3 (1-7)
Sampler euler
Scheduler simple
Base resolution 2048x2048
Default batch 1
Default upscaler None pinned -- falls back to 4x-AnimeSharp.pth

Good at: the longest and smallest text strings, fine detail at square-ish ratios, final renders where you already know the composition works.

Weak at: wide formats -- the 2048-per-side cap means 21:9 lands on exactly the same pixels as Turbo, so the extra 36 steps buy detail and nothing else; iteration, since batch drops to 1 and each render is 48 steps of a dual-transformer model on one home GPU; and it inherits every family limitation (no negatives, no Detailer, no USDU) while costing the most.

Ideogram 4 (Turbo)#

12 steps with a 1-step polish at a 1024x1024 base. The same weights as the default preset, sampled in a hurry.
Architecture Ideogram 4, 9.3B single-stream DiT (fp8) -- same weights as the default preset
Subtype none
Prompting Prose, short and concrete -- the enhancer rewrites it either way
Negatives Ignored
Steps 12 (12-48)
CFG 3 (1-7)
Sampler euler
Scheduler simple
Base resolution 1024x1024
Default batch 2
Default upscaler None pinned -- falls back to 4x-AnimeSharp.pth

Good at: seed hunting and checking whether the enhancer understood your idea before you spend 48 steps on it; still loads and runs the full dual-branch guidance, so composition matches what Quality will produce.

Weak at: fidelity, which its own config description says plainly -- 12 steps and a single polish step is where long strings, small type and fine detail fall apart first; it is also still the heavy dual-transformer load, so "Turbo" here means fewer steps, not a lighter model.

Turbo to find it, Quality to finish it

Because Turbo, the default preset and Quality are the same weights, a seed that composes well on Turbo composes the same way on Quality. Draft on Turbo, then re-roll the good seed at Quality.

Prompt enhancement is forced on#

Ideogram 4 always enhances, whatever your settings say

Every other family respects your AI-features setting. Ideogram 4 does not. Enhancement is turned on for you regardless of whether you have AI features enabled in /settings. The prompt you type is not the prompt that gets rendered.

This is deliberate. Ideogram 4 does not want a prompt -- it wants a structured caption, and the enhancer is what produces it. Your idea is passed through a vendor-supplied "magic prompt" system prompt of roughly 28 KB that rewrites it into a single line of minified JSON with exactly three keys:

Key What goes in it
high_level_description The whole scene in 50 words or fewer
style_description Aesthetics, lighting, medium, and either a photo or an art_style block, optionally up to five hex colours
compositional_deconstruction A background string plus a list of elements, each with a bounding box

The bot also prefixes your target aspect ratio onto the input so the caption is written for the shape you asked for.

What this costs you is control. Your wording, your emphasis and your ordering are all discarded; what survives is whatever the enhancer decided your scene contained. If the result is wrong, you cannot fix a word in your prompt and re-run -- you have to change your idea enough that the caption comes out differently.

You can turn it off with enhance: False or raw: True on /imagine. Do not, unless you are writing the JSON yourself. Output quality drops noticeably without it, because the model is expecting that shape of input.

What the magic prompt does to your idea#

Knowing the rules it follows makes it much easier to predict your result:

Rule Effect on your image
Photoreal by default Underspecified scenes default to an iPhone-snapshot aesthetic -- casual framing, ordinary light. Say "oil painting", "screen print", "3D render" explicitly if you want anything else.
The word "warm" is banned as a grading adjective It triggers the amber "AI look", so the enhancer will not use it. Naming a specific light source works better.
Off-centre framing preferred Perfectly centred compositions are avoided unless you ask for one.
No motion blur in candid photos Candid shots come out sharp.
One coherent subject is one element It will not split a person into head, torso and limbs, or a car into body and wheels.
Ground, floor, sky, crowd and horizon are always background This exists to stop figures being rendered with their legs sunk into the floor.
Every piece of text is its own element Each quoted string, sign, jersey number and price label is emitted separately, so you can ask for several distinct pieces of text and get them all.
Named things stay named Brands, public figures and characters are named explicitly rather than paraphrased.
No hedging Phrases like "things like", "various", "or similar" and "oak or walnut" are stripped. It always commits to something specific.

Write plainly and let it work

A short, concrete idea in ordinary English gives the enhancer room to build the caption. For text in the image, put the exact words in quotes: a neon sign reading "OPEN LATE". Quoted strings are the part of your input most likely to survive the rewrite intact.

Resolution limits#

Every side is clamped to 2048 pixels

Ideogram 4 hard-caps each dimension between 256 and 2048, in multiples of 16. Wide aspect ratios are produced by shrinking the short side, not by growing the long side.

The practical consequence: at 21:9 all three presets land on the same 2048-wide image, so Quality only buys you real resolution near square ratios. If you want a wide Ideogram image at more pixels, upscale it afterwards rather than asking for a bigger render.

What Ideogram 4 does not support#

Feature Status
Negative prompt Not supported. Anything in negative: is discarded on all four versions. Describe what you want instead.
Detailer Refused. Running it returns "Detailer is not compatible with ideogram4 models. Please use a standard model." There is no face, eye or hand fix pass for an Ideogram image -- if a face comes out wrong, your options are a re-roll or a different model.
USDU upscale Refused. You get "USDU upscale is not compatible with ideogram4 models. Try the regular upscale…"
Quality tags None shipped -- your caption is encoded as written
Safe-content guard strings None shipped, so your nsfw setting changes no prompt text here
Hi-res fix Supported -- no architecture is blocked from it
Model-based upscale and SeedVR2 Supported. Use these instead of USDU.

How to enlarge an Ideogram image

Since USDU is off the table, use /upscale with the Model-based (fast) method or SeedVR2. Both work on any image regardless of which model made it.

LoRAs and character names#

Two of the 91 LoRAs are Ideogram 4 compatible. Up to 4 stack at once. They patch the model only, never the text encoder, and on the three dual-branch versions they are applied to both guidance branches -- the unconditional branch gets about two thirds of the strength you set, which is what keeps the asymmetric guidance balanced. On Fast there is only one branch, so they stack there alone. See LoRAs.

Character names are written in natural form -- Hatsune Miku, not the Danbooru tag style used by SDXL. See Character matching.

Warning

Running Ideogram through /turbo silently drops any LoRAs you selected. They are logged as skipped, not reported on the card.

Categories: Models | AI Enhance | Imagine | Upscaling